Fix tg3__arena_strdup to distinguish empty strings from absent strings

Co-authored-by: syoyo <18676+syoyo@users.noreply.github.com>
Agent-Logs-Url: https://github.com/syoyo/tinygltf/sessions/445ab61b-4294-45e6-8faf-4f2fc8dfe369

.github/instructions/copilot-instructions.md

@@ -0,0 +1,92 @@
+# Copilot Review Instructions for TinyGLTF
+
+This document provides guidelines for reviewing code changes in the TinyGLTF repository.
+
+## Memory Safety
+
+- **Buffer Overflows**: Check for proper bounds checking when accessing arrays, vectors, and buffers. Verify that buffer sizes are validated before read/write operations.
+- **Null Pointer Dereferences**: Ensure all pointers are checked for null before dereferencing, especially when handling optional glTF fields.
+- **Memory Leaks**: Verify proper resource management, including RAII patterns for file handles, image data, and dynamically allocated memory.
+- **Use-After-Free**: Check for proper lifetime management of objects, especially when dealing with callbacks and asynchronous operations.
+
+## Error Handling
+
+- **File I/O**: Verify that all file operations have proper error checking and meaningful error messages.
+- **JSON Parsing**: Ensure JSON parsing errors are caught and reported with helpful context about the location and nature of the error.
+- **Resource Loading**: Check that failures in loading images, buffers, and other resources are properly handled and don't cause crashes.
+- **Error Propagation**: Verify that errors are properly propagated through the call stack with appropriate error messages.
+
+## glTF 2.0 Specification Compliance
+
+- **Required Fields**: Ensure all required glTF fields are validated and present.
+- **Data Types**: Verify that data types match the glTF specification (e.g., component types, accessor types).
+- **Constraints**: Check that glTF constraints are enforced (e.g., valid ranges for enums, buffer stride requirements).
+- **Extensions**: Verify proper handling of glTF extensions and that unknown extensions are handled gracefully.
+- **Validation**: Ensure new features align with the glTF 2.0 specification from the Khronos Group.
+
+## Cross-Platform Compatibility
+
+- **Windows**: Check for proper handling of Windows-specific issues (path separators, line endings, file operations).
+- **Linux**: Verify compatibility with various Linux distributions and compilers (GCC, Clang).
+- **macOS**: Ensure macOS-specific considerations are addressed (case-sensitive filesystems, Clang compatibility).
+- **Mobile Platforms**: Consider Android and iOS compatibility where applicable.
+- **Endianness**: Verify proper handling of byte order when reading binary data.
+- **Compiler Compatibility**: Ensure code compiles with C++11 standard and supported compilers (MSVC, GCC, Clang).
+
+## Edge Cases in glTF Parsing
+
+- **Empty/Minimal Files**: Verify handling of minimal valid glTF files.
+- **Large Files**: Check for proper handling of large glTF files and buffers without memory exhaustion.
+- **Malformed Data**: Ensure graceful handling of malformed or invalid glTF data.
+- **Missing Optional Fields**: Verify correct behavior when optional glTF fields are absent.
+- **Edge Values**: Check handling of boundary values (e.g., maximum buffer sizes, extreme floating-point values).
+- **Base64 Encoding**: Verify proper handling of base64-encoded data URIs and invalid encodings.
+
+## Backwards Compatibility
+
+- **API Changes**: Ensure public API changes maintain backwards compatibility or are properly deprecated.
+- **Breaking Changes**: Flag any breaking changes for major version updates and document migration paths.
+- **Binary Compatibility**: Consider ABI stability for header-only library changes.
+- **Default Behavior**: Verify that default behavior of existing functionality remains unchanged.
+
+## Performance Considerations
+
+- **Parsing Performance**: Check for unnecessary copies, redundant allocations, and inefficient algorithms in parsing logic.
+- **Memory Usage**: Verify efficient memory usage, especially when loading large glTF files.
+- **I/O Operations**: Ensure efficient file reading and minimize unnecessary disk access.
+- **String Operations**: Check for efficient string handling (use of string_view, move semantics).
+- **STL Usage**: Verify appropriate use of STL containers and algorithms.
+
+## Documentation
+
+- **Public API**: Ensure all public functions, classes, and methods have clear documentation comments.
+- **Parameters**: Verify that function parameters are documented, including expected ranges and constraints.
+- **Return Values**: Document return values and possible error conditions.
+- **Examples**: Check that complex features include usage examples.
+- **Changelog**: Verify that significant changes are documented in release notes or changelog.
+
+## Testing
+
+- **Test Coverage**: Ensure new features include appropriate unit tests or integration tests.
+- **Edge Cases**: Verify that tests cover edge cases and error conditions.
+- **Cross-Platform Tests**: Check that tests run on all supported platforms.
+- **Regression Tests**: Ensure bug fixes include regression tests to prevent recurrence.
+- **Sample Files**: Verify that changes are tested with various valid and invalid glTF sample files.
+
+## Code Style Consistency
+
+- **Header-Only Pattern**: Maintain the header-only library structure.
+- **Naming Conventions**: Follow existing naming conventions (CamelCase for types, snake_case for functions where applicable).
+- **Formatting**: Adhere to the existing code formatting style (check `.clang-format` if available).
+- **Include Guards**: Verify proper include guards and header organization.
+- **Namespace Usage**: Ensure proper use of the `tinygltf` namespace.
+- **Comments**: Maintain consistent comment style with existing code.
+- **C++11 Compliance**: Verify that code uses C++11 features appropriately and doesn't require newer standards unless specified.
+
+## Additional Considerations
+
+- **Third-Party Dependencies**: Minimize new dependencies; prefer existing dependencies (json.hpp, stb_image).
+- **Warnings**: Ensure code compiles without warnings on supported compilers.
+- **const Correctness**: Verify proper use of const for parameters and methods.
+- **RAII**: Prefer RAII patterns for resource management over manual cleanup.
+- **noexcept**: Use noexcept appropriately for move constructors and move assignment operators.

.github/workflows/ci.yml

@@ -0,0 +1,330 @@
+name: Comprehensive CI
+
+on:
+  push:
+    branches:
+      - master
+      - release
+      - devel
+  pull_request:
+    branches:
+      - master
+      - release
+      - devel
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  # Linux x64 - GCC
+  linux-gcc-x64:
+    runs-on: ubuntu-latest
+    name: Linux x64 (GCC)
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Configure
+        run: cmake -B build -DTINYGLTF_BUILD_LOADER_EXAMPLE=ON
+
+      - name: Build
+        run: cmake --build build
+
+      - name: Run loader_example
+        run: ./build/loader_example models/Cube/Cube.gltf
+
+      - name: Run tests
+        run: ctest --test-dir build --output-on-failure
+
+  # Linux x64 - Clang 21
+  linux-clang-x64:
+    runs-on: ubuntu-24.04
+    name: Linux x64 (Clang 21)
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Install Clang 21
+        run: |
+          wget https://apt.llvm.org/llvm.sh
+          chmod +x llvm.sh
+          sudo ./llvm.sh 21
+
+      - name: Configure
+        run: |
+          cmake -B build -DCMAKE_C_COMPILER=clang-21 -DCMAKE_CXX_COMPILER=clang++-21 -DTINYGLTF_BUILD_LOADER_EXAMPLE=ON
+
+      - name: Build
+        run: cmake --build build
+
+      - name: Run loader_example
+        run: |
+          ./build/loader_example models/Cube/Cube.gltf
+
+      - name: Run tests
+        run: ctest --test-dir build --output-on-failure
+
+  # Linux ARM64 - GCC (native)
+  linux-arm64:
+    runs-on: ubuntu-24.04-arm
+    name: Linux ARM64 (GCC)
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Configure
+        run: cmake -B build -DTINYGLTF_BUILD_LOADER_EXAMPLE=ON
+
+      - name: Build
+        run: cmake --build build
+
+      - name: Run loader_example
+        run: ./build/loader_example models/Cube/Cube.gltf
+
+      - name: Run tests
+        run: ctest --test-dir build --output-on-failure
+
+  # macOS ARM64 Apple Silicon
+  macos-arm64:
+    runs-on: macos-14
+    name: macOS ARM64 Apple Silicon (Clang)
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Configure
+        run: cmake -B build -DTINYGLTF_BUILD_LOADER_EXAMPLE=ON
+
+      - name: Build
+        run: cmake --build build
+
+      - name: Run loader_example
+        run: ./build/loader_example models/Cube/Cube.gltf
+
+      - name: Run tests
+        run: ctest --test-dir build --output-on-failure
+
+  # Windows x64 - MSVC
+  windows-msvc-x64:
+    runs-on: windows-latest
+    name: Windows x64 (MSVC)
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Configure
+        run: |
+          mkdir build
+          cd build
+          cmake -G "Visual Studio 17 2022" -A x64 -DTINYGLTF_BUILD_LOADER_EXAMPLE=On -DTINYGLTF_BUILD_GL_EXAMPLES=Off -DTINYGLTF_BUILD_VALIDATOR_EXAMPLE=Off ..
+
+      - name: Build
+        run: cmake --build build --config Release
+
+      - name: Run loader_example
+        run: |
+          .\build\Release\loader_example.exe models\Cube\Cube.gltf
+
+      - name: Run tests
+        run: ctest --test-dir build -C Release --output-on-failure
+
+  # Windows x86 - MSVC
+  windows-msvc-x86:
+    runs-on: windows-latest
+    name: Windows x86 (MSVC)
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Configure
+        run: |
+          mkdir build
+          cd build
+          cmake -G "Visual Studio 17 2022" -A Win32 -DTINYGLTF_BUILD_LOADER_EXAMPLE=On -DTINYGLTF_BUILD_GL_EXAMPLES=Off -DTINYGLTF_BUILD_VALIDATOR_EXAMPLE=Off ..
+
+      - name: Build
+        run: cmake --build build --config Release
+
+      - name: Run tests
+        run: ctest --test-dir build -C Release --output-on-failure
+
+  # Windows ARM64 - MSVC (cross-compile)
+  windows-msvc-arm64:
+    runs-on: windows-latest
+    name: Windows ARM64 (MSVC) - Cross-compile
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Configure
+        run: |
+          mkdir build
+          cd build
+          cmake -G "Visual Studio 17 2022" -A ARM64 -DTINYGLTF_BUILD_LOADER_EXAMPLE=On -DTINYGLTF_BUILD_GL_EXAMPLES=Off -DTINYGLTF_BUILD_VALIDATOR_EXAMPLE=Off ..
+
+      - name: Build
+        run: cmake --build build --config Release
+
+  # Windows MinGW - MSYS2
+  windows-mingw-msys2:
+    runs-on: windows-latest
+    name: Windows x64 (MinGW MSYS2)
+    defaults:
+      run:
+        shell: msys2 {0}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Setup MSYS2
+        uses: msys2/setup-msys2@v2
+        with:
+          msystem: UCRT64
+          install: base-devel
+          pacboy: >-
+            cc:p cmake:p ninja:p
+          update: true
+          release: false
+
+      - name: Build with CMake
+        run: |
+          cmake -G"Ninja" -S . -B build
+          cmake --build build
+
+      - name: Run loader_example
+        run: |
+          ./build/loader_example models/Cube/Cube.gltf
+
+      - name: Run tests
+        run: ctest --test-dir build --output-on-failure
+
+  # Linux -> Windows MinGW Cross-compile
+  linux-mingw-cross:
+    runs-on: ubuntu-latest
+    name: Linux→Windows (MinGW Cross) - Build Only
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Install MinGW
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y build-essential mingw-w64
+
+      - name: Build
+        run: |
+          x86_64-w64-mingw32-g++ -std=c++11 -o loader_example.exe loader_example.cc
+
+  # Special Configuration: No Exceptions
+  linux-noexception:
+    runs-on: ubuntu-latest
+    name: Linux x64 (GCC) - No Exceptions
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Build loader_example
+        run: |
+          g++ -DTINYGLTF_NOEXCEPTION -std=c++11 -o loader_example loader_example.cc
+
+      - name: Run loader_example
+        run: |
+          ./loader_example models/Cube/Cube.gltf
+
+      - name: Build and run unit tests
+        run: |
+          cd tests
+          g++ -DTINYGLTF_NOEXCEPTION -I../ -std=c++11 -g -O0 -o tester_noexcept tester.cc
+          ./tester_noexcept
+
+  # Special Configuration: Header-Only Mode
+  linux-header-only:
+    runs-on: ubuntu-latest
+    name: Linux x64 (GCC) - Header-Only Mode
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Build with CMake Header-Only
+        run: |
+          mkdir build
+          cmake -B build -DTINYGLTF_HEADER_ONLY=ON -DTINYGLTF_BUILD_LOADER_EXAMPLE=ON
+          cmake --build build
+
+      - name: Run loader_example
+        run: |
+          ./build/loader_example models/Cube/Cube.gltf
+
+      - name: Run tests
+        run: ctest --test-dir build --output-on-failure
+
+  # Special Configuration: RapidJSON Backend
+  linux-rapidjson:
+    runs-on: ubuntu-latest
+    name: Linux x64 (GCC) - RapidJSON Backend
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Clone RapidJSON
+        run: |
+          git clone https://github.com/Tencent/rapidjson
+
+      - name: Configure
+        run: |
+          cmake -B build -DTINYGLTF_USE_RAPIDJSON=ON -DTINYGLTF_BUILD_LOADER_EXAMPLE=ON -DCMAKE_PREFIX_PATH=$PWD/rapidjson
+
+      - name: Build
+        run: cmake --build build
+
+      - name: Run loader_example
+        run: |
+          ./build/loader_example models/Cube/Cube.gltf
+
+      - name: Run tests
+        run: ctest --test-dir build --output-on-failure
+
+  # Special Configuration: AddressSanitizer
+  linux-asan:
+    runs-on: ubuntu-latest
+    name: Linux x64 (Clang) - AddressSanitizer
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Build loader_example with ASan
+        run: |
+          clang++ -fsanitize=address -std=c++11 -g -O1 -o loader_example loader_example.cc
+
+      - name: Run loader_example
+        run: |
+          ./loader_example models/Cube/Cube.gltf
+
+      - name: Build and run unit tests with ASan
+        run: |
+          cd tests
+          clang++ -fsanitize=address -I../ -std=c++11 -g -O1 -o tester tester.cc
+          ./tester
+
+  # Special Configuration: UndefinedBehaviorSanitizer
+  linux-ubsan:
+    runs-on: ubuntu-latest
+    name: Linux x64 (Clang) - UndefinedBehaviorSanitizer
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Build loader_example with UBSan
+        run: |
+          clang++ -fsanitize=undefined -std=c++11 -g -O1 -o loader_example loader_example.cc
+
+      - name: Run loader_example
+        run: |
+          ./loader_example models/Cube/Cube.gltf
+
+      - name: Build and run unit tests with UBSan
+        run: |
+          cd tests
+          clang++ -fsanitize=undefined -I../ -std=c++11 -g -O1 -o tester tester.cc
+          ./tester

.gitignore

@@ -21,6 +21,9 @@ premake5.tar.gz
 *.vcxproj*
 .vs
 
+# default cmake build dir
+build/
+
 #binary directories
 bin/
 obj/
@@ -70,6 +73,19 @@ tests/tester
 tests/tester_noexcept
 tests/issue-97.gltf
 tests/issue-261.gltf
+tests/issue-495-external.gltf
+# Test-generated output files (written by tester.cc during test run)
+tests/Cube.gltf
+tests/Cube.bin
+tests/Cube.glb
+tests/Cube_BaseColor.png
+tests/Cube_MetallicRoughness.png
+tests/Cube_with_embedded_images.gltf
+tests/Cube_with_image_files.gltf
+tests/tmp.glb
+tests/ issue-236.gltf
+tests/ issue-236.bin
+tests/ 2x2 image  has multiple      spaces.png
 
 # unignore
 !Makefile

CMakeLists.txt

@@ -14,9 +14,11 @@ option(TINYGLTF_BUILD_LOADER_EXAMPLE "Build loader_example(load glTF and dump in
 option(TINYGLTF_BUILD_GL_EXAMPLES "Build GL exampels(requires glfw, OpenGL, etc)" OFF)
 option(TINYGLTF_BUILD_VALIDATOR_EXAMPLE "Build validator exampe" OFF)
 option(TINYGLTF_BUILD_BUILDER_EXAMPLE "Build glTF builder example" OFF)
+option(TINYGLTF_BUILD_TESTS "Build unit tests" OFF)
 option(TINYGLTF_HEADER_ONLY "On: header-only mode. Off: create tinygltf library(No TINYGLTF_IMPLEMENTATION required in your project)" OFF)
 option(TINYGLTF_INSTALL "Install tinygltf files during install step. Usually set to OFF if you include tinygltf through add_subdirectory()" ON)
 option(TINYGLTF_INSTALL_VENDOR "Install vendored nlohmann/json and nothings/stb headers" ON)
+option(TINYGLTF_USE_CUSTOM_JSON "Use the built-in fast JSON parser (tinygltf_json.h) instead of nlohmann/json" OFF)
 
 if (TINYGLTF_BUILD_LOADER_EXAMPLE)
   add_executable(loader_example
@@ -37,6 +39,25 @@ if (TINYGLTF_BUILD_BUILDER_EXAMPLE)
   add_subdirectory ( examples/build-gltf )
 endif (TINYGLTF_BUILD_BUILDER_EXAMPLE)
 
+if (TINYGLTF_BUILD_TESTS)
+  enable_testing()
+  add_executable(tester tests/tester.cc)
+  target_include_directories(tester PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}
+    ${CMAKE_CURRENT_SOURCE_DIR}/tests
+  )
+  add_test(NAME tester COMMAND tester WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/tests)
+
+  # Build and run tests with the custom JSON backend enabled to catch regressions
+  add_executable(tester_customjson tests/tester.cc)
+  target_include_directories(tester_customjson PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}
+    ${CMAKE_CURRENT_SOURCE_DIR}/tests
+  )
+  target_compile_definitions(tester_customjson PRIVATE TINYGLTF_USE_CUSTOM_JSON)
+  add_test(NAME tester_customjson COMMAND tester_customjson WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/tests)
+endif (TINYGLTF_BUILD_TESTS)
+
 #
 # for add_subdirectory and standalone build
 #
@@ -60,6 +81,14 @@ else (TINYGLTF_HEADER_ONLY)
           )
 endif (TINYGLTF_HEADER_ONLY)
 
+if (TINYGLTF_USE_CUSTOM_JSON)
+  if (TINYGLTF_HEADER_ONLY)
+    target_compile_definitions(tinygltf INTERFACE TINYGLTF_USE_CUSTOM_JSON)
+  else ()
+    target_compile_definitions(tinygltf PUBLIC TINYGLTF_USE_CUSTOM_JSON)
+  endif ()
+endif ()
+
 if (TINYGLTF_INSTALL)
   install(TARGETS tinygltf EXPORT tinygltfTargets)
   install(EXPORT tinygltfTargets NAMESPACE tinygltf:: FILE TinyGLTFTargets.cmake DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/tinygltf)
@@ -69,6 +98,7 @@ if (TINYGLTF_INSTALL)
 
   INSTALL ( FILES
     tiny_gltf.h
+    tinygltf_json.h
     ${TINYGLTF_EXTRA_SOUECES}
     DESTINATION
     include

README.md

@@ -26,8 +26,6 @@ Currently TinyGLTF is stable and maintenance mode. No drastic changes and featur
 
 ## Builds
 
-[![Build status](https://ci.appveyor.com/api/projects/status/warngenu9wjjhlm8?svg=true)](https://ci.appveyor.com/project/syoyo/tinygltf)
-
 ![C/C++ CI](https://github.com/syoyo/tinygltf/workflows/C/C++%20CI/badge.svg)
 
 ## Features
@@ -159,9 +157,10 @@ Model model;
 TinyGLTF loader;
 std::string err;
 std::string warn;
+std::string filename = "input.gltf";
 
-bool ret = loader.LoadASCIIFromFile(&model, &err, &warn, argv[1]);
-//bool ret = loader.LoadBinaryFromFile(&model, &err, &warn, argv[1]); // for binary glTF(.glb)
+bool ret = loader.LoadASCIIFromFile(&model, &err, &warn, filename);
+//bool ret = loader.LoadBinaryFromFile(&model, &err, &warn, filename); // for binary glTF(.glb)
 
 if (!warn.empty()) {
   printf("Warn: %s\n", warn.c_str());
@@ -172,8 +171,7 @@ if (!err.empty()) {
 }
 
 if (!ret) {
-  printf("Failed to parse glTF\n");
-  return -1;
+  printf("Failed to parse glTF: %s\n", filename.c_str());
 }
 ```
 
@@ -194,7 +192,6 @@ if (!ret) {
 * `TINYGLTF_NO_INCLUDE_STB_IMAGE `: Disable including `stb_image.h` from within `tiny_gltf.h` because it has been already included before or you want to include it using custom path before including `tiny_gltf.h`.
 * `TINYGLTF_NO_INCLUDE_STB_IMAGE_WRITE `: Disable including `stb_image_write.h` from within `tiny_gltf.h` because it has been already included before or you want to include it using custom path before including `tiny_gltf.h`.
 * `TINYGLTF_USE_RAPIDJSON` : Use RapidJSON as a JSON parser/serializer. RapidJSON files are not included in TinyGLTF repo. Please set an include path to RapidJSON if you enable this feature.
-* `TINYGLTF_USE_CPP14` : Use C++14 feature(requires C++14 compiler). This may give better performance than C++11.
 
 
 ## CMake options

appveyor.yml

@@ -1,18 +0,0 @@
-version: 0.9.{build}
-
-image:
-  - Visual Studio 2015
-
-# scripts that runs after repo cloning.
-install:
-  - vcsetup.bat
-
-platform: x64
-configuration: Release
-
-build:
-  parallel: true
-  project: TinyGLTFSolution.sln
-
-after_build:
-  - examples.bat

benchmark/Makefile

@@ -0,0 +1,70 @@
+# benchmark/Makefile — Build and run tinygltf v3 benchmarks
+#
+# Targets:
+#   make           — build gen_synthetic + bench_v3
+#   make generate  — generate synthetic test scenes
+#   make run       — run benchmarks on all generated scenes
+#   make report    — run benchmarks and produce CSV report
+#   make clean     — remove binaries and generated scenes
+
+CXX       ?= g++
+CXXFLAGS  ?= -O2 -std=c++17 -Wall -Wextra -Wno-unused-function
+CXXFLAGS  += -fno-rtti -fno-exceptions
+INCLUDES  = -I..
+
+BINDIR    = .
+GEN       = $(BINDIR)/gen_synthetic
+BENCH_V3  = $(BINDIR)/bench_v3
+
+# Iteration counts
+ITERATIONS ?= 10
+WARMUP     ?= 2
+PREFIX     ?= synthetic
+
+.PHONY: all generate run report clean
+
+all: $(GEN) $(BENCH_V3)
+
+$(GEN): gen_synthetic.cpp
+	$(CXX) $(CXXFLAGS) -o $@ $<
+
+$(BENCH_V3): bench_v3.cpp ../tiny_gltf_v3.h ../tinygltf_json.h
+	$(CXX) $(CXXFLAGS) $(INCLUDES) -o $@ $<
+
+# Generate synthetic scenes of varying sizes
+generate: $(GEN)
+	@echo "=== Generating synthetic scenes ==="
+	./$(GEN) --prefix $(PREFIX)
+	@echo ""
+	@echo "Generated files (binary + GLB):"
+	@ls -lh $(PREFIX)_*.gltf $(PREFIX)_*.glb $(PREFIX)_*.bin 2>/dev/null || true
+
+# Run benchmarks on all generated scenes
+run: $(BENCH_V3) generate
+	@echo ""
+	@echo "================================================================="
+	@echo "  tinygltf v3 Benchmark"
+	@echo "================================================================="
+	@echo ""
+	@for f in $(PREFIX)_*.glb $(PREFIX)_*.gltf; do \
+		if [ -f "$$f" ]; then \
+			./$(BENCH_V3) "$$f" --iterations $(ITERATIONS) --warmup $(WARMUP); \
+			echo ""; \
+		fi; \
+	done
+
+# Run benchmarks and produce CSV report
+report: $(BENCH_V3) generate
+	@echo "file,size_bytes,iterations,parse_min_ms,parse_max_ms,parse_avg_ms,parse_median_ms,throughput_mbs,arena_peak_bytes,meshes,nodes,accessors,materials,animations" > benchmark_report.csv
+	@for f in $(PREFIX)_*.glb $(PREFIX)_*.gltf; do \
+		if [ -f "$$f" ]; then \
+			./$(BENCH_V3) "$$f" --iterations $(ITERATIONS) --warmup $(WARMUP) --csv | tail -1 >> benchmark_report.csv; \
+		fi; \
+	done
+	@echo "=== Report written to benchmark_report.csv ==="
+	@cat benchmark_report.csv | column -t -s,
+
+clean:
+	rm -f $(GEN) $(BENCH_V3)
+	rm -f $(PREFIX)_*.gltf $(PREFIX)_*.glb $(PREFIX)_*.bin
+	rm -f benchmark_report.csv

benchmark/bench_v3.cpp

@@ -0,0 +1,396 @@
+/*
+ * bench_v3.cpp — Benchmark tinygltf v3 parser: parse speed & memory.
+ *
+ * Measures:
+ *   - File read time
+ *   - JSON parse + model build time
+ *   - Peak arena memory usage
+ *   - Throughput (MB/s)
+ *
+ * Usage:
+ *   bench_v3 <file.gltf|file.glb> [--iterations N] [--warmup N] [--quiet]
+ *   bench_v3 --batch <file1> <file2> ... [--iterations N]
+ */
+
+#define TINYGLTF3_IMPLEMENTATION
+#define TINYGLTF3_ENABLE_FS
+#include "tiny_gltf_v3.h"
+
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <cmath>
+#include <vector>
+#include <string>
+#include <algorithm>
+#include <chrono>
+
+#if defined(__linux__)
+#include <sys/resource.h>
+#endif
+
+/* ------------------------------------------------------------------ */
+/* Timing helpers                                                     */
+/* ------------------------------------------------------------------ */
+
+using Clock = std::chrono::high_resolution_clock;
+using TimePoint = Clock::time_point;
+
+static double elapsed_ms(TimePoint start, TimePoint end) {
+    return std::chrono::duration<double, std::milli>(end - start).count();
+}
+
+/* ------------------------------------------------------------------ */
+/* Memory tracking allocator                                          */
+/* ------------------------------------------------------------------ */
+
+struct MemTracker {
+    size_t current;
+    size_t peak;
+    size_t total_allocs;
+    size_t total_frees;
+};
+
+static void *tracked_alloc(size_t size, void *ud) {
+    MemTracker *mt = (MemTracker *)ud;
+    void *ptr = malloc(size);
+    if (ptr) {
+        mt->current += size;
+        if (mt->current > mt->peak) mt->peak = mt->current;
+        mt->total_allocs++;
+    }
+    return ptr;
+}
+
+static void *tracked_realloc(void *ptr, size_t old_size, size_t new_size, void *ud) {
+    MemTracker *mt = (MemTracker *)ud;
+    void *new_ptr = realloc(ptr, new_size);
+    if (new_ptr) {
+        mt->current -= old_size;
+        mt->current += new_size;
+        if (mt->current > mt->peak) mt->peak = mt->current;
+    }
+    return new_ptr;
+}
+
+static void tracked_free(void *ptr, size_t size, void *ud) {
+    MemTracker *mt = (MemTracker *)ud;
+    if (ptr) {
+        mt->current -= size;
+        mt->total_frees++;
+        free(ptr);
+    }
+}
+
+/* ------------------------------------------------------------------ */
+/* RSS measurement (Linux)                                            */
+/* ------------------------------------------------------------------ */
+
+static size_t get_rss_bytes() {
+#if defined(__linux__)
+    FILE *f = fopen("/proc/self/statm", "r");
+    if (!f) return 0;
+    long pages = 0;
+    if (fscanf(f, "%*s %ld", &pages) != 1) pages = 0;
+    fclose(f);
+    return (size_t)pages * 4096;
+#else
+    return 0;
+#endif
+}
+
+/* ------------------------------------------------------------------ */
+/* Benchmark result                                                   */
+/* ------------------------------------------------------------------ */
+
+struct BenchResult {
+    std::string filename;
+    uint64_t    file_size;
+    int         iterations;
+
+    /* Parse timing (ms) */
+    double parse_min;
+    double parse_max;
+    double parse_avg;
+    double parse_median;
+
+    /* Memory */
+    size_t arena_peak;     /* Peak arena allocation */
+    size_t rss_before;
+    size_t rss_after;
+
+    /* Model stats */
+    uint32_t meshes;
+    uint32_t nodes;
+    uint32_t accessors;
+    uint32_t materials;
+    uint32_t animations;
+    uint32_t buffers;
+    uint32_t buffer_views;
+    uint32_t images;
+    uint32_t textures;
+
+    /* Derived */
+    double throughput_mbs; /* MB/s based on median */
+};
+
+/* ------------------------------------------------------------------ */
+/* Run benchmark for a single file                                    */
+/* ------------------------------------------------------------------ */
+
+static BenchResult bench_file(const char *filename, int iterations, int warmup,
+                               bool quiet, int float32_mode = 0) {
+    BenchResult r = {};
+    r.filename = filename;
+    r.iterations = iterations;
+
+    /* Read file into memory */
+    FILE *f = fopen(filename, "rb");
+    if (!f) {
+        fprintf(stderr, "ERROR: Cannot open '%s'\n", filename);
+        return r;
+    }
+    fseek(f, 0, SEEK_END);
+    long sz = ftell(f);
+    fseek(f, 0, SEEK_SET);
+    if (sz <= 0) { fclose(f); return r; }
+
+    std::vector<uint8_t> data((size_t)sz);
+    size_t rd = fread(data.data(), 1, (size_t)sz, f);
+    fclose(f);
+    if ((long)rd != sz) { return r; }
+
+    r.file_size = (uint64_t)sz;
+
+    /* Extract base dir */
+    std::string path(filename);
+    std::string base_dir;
+    size_t sep = path.find_last_of("/\\");
+    if (sep != std::string::npos) base_dir = path.substr(0, sep);
+
+    /* Warmup iterations (not measured) */
+    for (int i = 0; i < warmup; ++i) {
+        tg3_model model;
+        tg3_error_stack errors;
+        tg3_error_stack_init(&errors);
+        tg3_parse_auto(&model, &errors, data.data(), data.size(),
+                       base_dir.c_str(), (uint32_t)base_dir.size(), NULL);
+        tg3_model_free(&model);
+        tg3_error_stack_free(&errors);
+    }
+
+    /* Benchmark iterations */
+    std::vector<double> times;
+    times.reserve(iterations);
+
+    MemTracker tracker_best;
+    memset(&tracker_best, 0, sizeof(tracker_best));
+
+    r.rss_before = get_rss_bytes();
+
+    for (int i = 0; i < iterations; ++i) {
+        MemTracker tracker;
+        memset(&tracker, 0, sizeof(tracker));
+
+        tg3_parse_options opts;
+        tg3_parse_options_init(&opts);
+        opts.memory.allocator.alloc = tracked_alloc;
+        opts.memory.allocator.realloc = tracked_realloc;
+        opts.memory.allocator.free = tracked_free;
+        opts.memory.allocator.user_data = &tracker;
+        opts.parse_float32 = float32_mode;
+
+        tg3_model model;
+        tg3_error_stack errors;
+        tg3_error_stack_init(&errors);
+
+        TimePoint t0 = Clock::now();
+
+        tg3_error_code err = tg3_parse_auto(&model, &errors,
+                                             data.data(), data.size(),
+                                             base_dir.c_str(),
+                                             (uint32_t)base_dir.size(),
+                                             &opts);
+
+        TimePoint t1 = Clock::now();
+        double ms = elapsed_ms(t0, t1);
+        times.push_back(ms);
+
+        /* Capture model stats on first successful iteration */
+        if (i == 0 && err == TG3_OK) {
+            r.meshes = model.meshes_count;
+            r.nodes = model.nodes_count;
+            r.accessors = model.accessors_count;
+            r.materials = model.materials_count;
+            r.animations = model.animations_count;
+            r.buffers = model.buffers_count;
+            r.buffer_views = model.buffer_views_count;
+            r.images = model.images_count;
+            r.textures = model.textures_count;
+        }
+
+        if (tracker.peak > tracker_best.peak) {
+            tracker_best = tracker;
+        }
+
+        tg3_model_free(&model);
+        tg3_error_stack_free(&errors);
+
+        if (err != TG3_OK && !quiet) {
+            fprintf(stderr, "  Parse error on iteration %d: code %d\n", i, (int)err);
+        }
+    }
+
+    r.rss_after = get_rss_bytes();
+    r.arena_peak = tracker_best.peak;
+
+    /* Compute stats */
+    std::sort(times.begin(), times.end());
+    r.parse_min = times.front();
+    r.parse_max = times.back();
+    double sum = 0;
+    for (double t : times) sum += t;
+    r.parse_avg = sum / times.size();
+    r.parse_median = times[times.size() / 2];
+
+    /* Throughput: file_size / median_time */
+    if (r.parse_median > 0) {
+        r.throughput_mbs = ((double)r.file_size / (1024.0 * 1024.0)) /
+                           (r.parse_median / 1000.0);
+    }
+
+    return r;
+}
+
+/* ------------------------------------------------------------------ */
+/* Print results                                                      */
+/* ------------------------------------------------------------------ */
+
+static const char *human_bytes(size_t bytes, char *buf, size_t buf_sz) {
+    if (bytes >= 1024ULL * 1024 * 1024)
+        snprintf(buf, buf_sz, "%.2f GB", (double)bytes / (1024.0 * 1024 * 1024));
+    else if (bytes >= 1024 * 1024)
+        snprintf(buf, buf_sz, "%.2f MB", (double)bytes / (1024.0 * 1024));
+    else if (bytes >= 1024)
+        snprintf(buf, buf_sz, "%.2f KB", (double)bytes / 1024.0);
+    else
+        snprintf(buf, buf_sz, "%zu B", bytes);
+    return buf;
+}
+
+static void print_result(const BenchResult &r) {
+    char buf1[64], buf2[64];
+
+    printf("┌─────────────────────────────────────────────────────────────────┐\n");
+    printf("│ %-63s │\n", r.filename.c_str());
+    printf("├─────────────────────────────────────────────────────────────────┤\n");
+    printf("│ File size:      %-47s │\n", human_bytes((size_t)r.file_size, buf1, sizeof(buf1)));
+    printf("│ Iterations:     %-47d │\n", r.iterations);
+    printf("│                                                                 │\n");
+    printf("│ Parse time (ms):                                                │\n");
+    printf("│   min:    %10.3f                                            │\n", r.parse_min);
+    printf("│   max:    %10.3f                                            │\n", r.parse_max);
+    printf("│   avg:    %10.3f                                            │\n", r.parse_avg);
+    printf("│   median: %10.3f                                            │\n", r.parse_median);
+    printf("│                                                                 │\n");
+    printf("│ Throughput:     %-47s │\n",
+           (snprintf(buf1, sizeof(buf1), "%.2f MB/s", r.throughput_mbs), buf1));
+    printf("│ Arena peak:     %-47s │\n", human_bytes(r.arena_peak, buf1, sizeof(buf1)));
+    if (r.rss_before > 0) {
+        printf("│ RSS before:     %-47s │\n", human_bytes(r.rss_before, buf1, sizeof(buf1)));
+        printf("│ RSS after:      %-47s │\n", human_bytes(r.rss_after, buf2, sizeof(buf2)));
+    }
+    printf("│                                                                 │\n");
+    printf("│ Model: %u meshes, %u nodes, %u accessors, %u materials",
+           r.meshes, r.nodes, r.accessors, r.materials);
+    printf("        │\n");
+    printf("│        %u animations, %u buffers, %u images",
+           r.animations, r.buffers, r.images);
+    printf("                       │\n");
+    printf("└─────────────────────────────────────────────────────────────────┘\n");
+}
+
+static void print_csv_header() {
+    printf("file,size_bytes,iterations,parse_min_ms,parse_max_ms,parse_avg_ms,"
+           "parse_median_ms,throughput_mbs,arena_peak_bytes,"
+           "meshes,nodes,accessors,materials,animations\n");
+}
+
+static void print_csv_row(const BenchResult &r) {
+    printf("%s,%lu,%d,%.3f,%.3f,%.3f,%.3f,%.2f,%zu,%u,%u,%u,%u,%u\n",
+           r.filename.c_str(), (unsigned long)r.file_size, r.iterations,
+           r.parse_min, r.parse_max, r.parse_avg, r.parse_median,
+           r.throughput_mbs, r.arena_peak,
+           r.meshes, r.nodes, r.accessors, r.materials, r.animations);
+}
+
+/* ------------------------------------------------------------------ */
+/* Main                                                               */
+/* ------------------------------------------------------------------ */
+
+static void usage() {
+    fprintf(stderr,
+        "Usage:\n"
+        "  bench_v3 <file> [--iterations N] [--warmup N] [--csv] [--quiet]\n"
+        "  bench_v3 --batch <file1> [file2] ... [--iterations N] [--csv]\n"
+        "\n"
+        "Options:\n"
+        "  --iterations N   Number of timed parse iterations (default: 10)\n"
+        "  --warmup N       Number of warmup iterations (default: 2)\n"
+        "  --csv            Output in CSV format\n"
+        "  --quiet          Suppress per-iteration error messages\n"
+        "  --batch          Benchmark multiple files\n"
+        "  --float32        Parse JSON floats as float32 (faster, less precise)\n");
+}
+
+int main(int argc, char **argv) {
+    if (argc < 2) { usage(); return 1; }
+
+    int iterations = 10;
+    int warmup = 2;
+    bool csv = false;
+    bool quiet = false;
+    int float32_mode = 0;
+    std::vector<std::string> files;
+
+    for (int i = 1; i < argc; ++i) {
+        if (strcmp(argv[i], "--iterations") == 0 && i + 1 < argc) {
+            iterations = atoi(argv[++i]);
+        } else if (strcmp(argv[i], "--warmup") == 0 && i + 1 < argc) {
+            warmup = atoi(argv[++i]);
+        } else if (strcmp(argv[i], "--csv") == 0) {
+            csv = true;
+        } else if (strcmp(argv[i], "--quiet") == 0) {
+            quiet = true;
+        } else if (strcmp(argv[i], "--float32") == 0) {
+            float32_mode = 1;
+        } else if (strcmp(argv[i], "--batch") == 0) {
+            /* batch mode: just collect files */
+        } else if (argv[i][0] != '-') {
+            files.push_back(argv[i]);
+        }
+    }
+
+    if (files.empty()) { usage(); return 1; }
+
+    if (csv) print_csv_header();
+
+    for (const auto &file : files) {
+        if (!csv && !quiet) {
+            printf("Benchmarking: %s (%d iterations, %d warmup%s)\n",
+                   file.c_str(), iterations, warmup,
+                   float32_mode ? ", float32" : "");
+        }
+
+        BenchResult r = bench_file(file.c_str(), iterations, warmup, quiet, float32_mode);
+
+        if (csv) {
+            print_csv_row(r);
+        } else {
+            print_result(r);
+            printf("\n");
+        }
+    }
+
+    return 0;
+}

benchmark/gen_synthetic.cpp

@@ -0,0 +1,740 @@
+/*
+ * gen_synthetic.cpp — Generate synthetic glTF 2.0 scenes for benchmarking.
+ *
+ * Produces .gltf (ASCII) and .glb (binary) files with configurable:
+ *   - Number of meshes, each with N vertices/triangles
+ *   - Number of nodes (flat hierarchy)
+ *   - Number of materials
+ *   - Number of animations with M keyframes
+ *
+ * Usage:
+ *   gen_synthetic [--meshes N] [--verts-per-mesh N] [--nodes N]
+ *                 [--materials N] [--animations N] [--keyframes N]
+ *                 [--prefix NAME]
+ *
+ * Outputs: <prefix>_<label>.gltf and <prefix>_<label>.glb
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <cstdint>
+
+/* ------------------------------------------------------------------ */
+/* Tiny JSON writer (no dependencies)                                 */
+/* ------------------------------------------------------------------ */
+
+struct JsonWriter {
+    std::string buf;
+    int indent;
+    bool need_comma;
+    std::vector<bool> stack; /* true = array context */
+
+    JsonWriter() : indent(0), need_comma(false) {}
+
+    void comma() {
+        if (need_comma) buf += ",";
+        buf += "\n";
+        for (int i = 0; i < indent; ++i) buf += "  ";
+    }
+
+    void begin_obj() {
+        if (!stack.empty()) comma();
+        buf += "{";
+        indent++;
+        need_comma = false;
+        stack.push_back(false);
+    }
+    void end_obj() {
+        indent--;
+        buf += "\n";
+        for (int i = 0; i < indent; ++i) buf += "  ";
+        buf += "}";
+        stack.pop_back();
+        need_comma = true;
+    }
+
+    void begin_arr() {
+        if (!stack.empty() && !need_comma) { /* first elem */ }
+        buf += "[";
+        indent++;
+        need_comma = false;
+        stack.push_back(true);
+    }
+    void end_arr() {
+        indent--;
+        buf += "\n";
+        for (int i = 0; i < indent; ++i) buf += "  ";
+        buf += "]";
+        stack.pop_back();
+        need_comma = true;
+    }
+
+    void key(const char *k) {
+        comma();
+        buf += "\"";
+        buf += k;
+        buf += "\": ";
+        need_comma = false;
+    }
+
+    void val_str(const char *v) {
+        if (stack.back()) comma();
+        buf += "\"";
+        buf += v;
+        buf += "\"";
+        need_comma = true;
+    }
+    void val_int(int64_t v) {
+        if (stack.back()) comma();
+        buf += std::to_string(v);
+        need_comma = true;
+    }
+    void val_double(double v) {
+        if (stack.back()) comma();
+        char tmp[64];
+        snprintf(tmp, sizeof(tmp), "%.6g", v);
+        buf += tmp;
+        need_comma = true;
+    }
+    void val_bool(bool v) {
+        if (stack.back()) comma();
+        buf += v ? "true" : "false";
+        need_comma = true;
+    }
+
+    void kv_str(const char *k, const char *v) { key(k); val_str(v); need_comma = true; }
+    void kv_int(const char *k, int64_t v) { key(k); val_int(v); need_comma = true; }
+    void kv_double(const char *k, double v) { key(k); val_double(v); need_comma = true; }
+    void kv_bool(const char *k, bool v) { key(k); val_bool(v); need_comma = true; }
+};
+
+/* ------------------------------------------------------------------ */
+/* Binary buffer builder                                              */
+/* ------------------------------------------------------------------ */
+
+struct BinBuffer {
+    std::vector<uint8_t> data;
+
+    size_t offset() const { return data.size(); }
+
+    void push_float(float v) {
+        const uint8_t *p = reinterpret_cast<const uint8_t*>(&v);
+        data.insert(data.end(), p, p + 4);
+    }
+    void push_u16(uint16_t v) {
+        const uint8_t *p = reinterpret_cast<const uint8_t*>(&v);
+        data.insert(data.end(), p, p + 2);
+    }
+    void push_u32(uint32_t v) {
+        const uint8_t *p = reinterpret_cast<const uint8_t*>(&v);
+        data.insert(data.end(), p, p + 4);
+    }
+    void align4() {
+        while (data.size() % 4 != 0) data.push_back(0);
+    }
+};
+
+/* ------------------------------------------------------------------ */
+/* Scene config                                                       */
+/* ------------------------------------------------------------------ */
+
+struct SceneConfig {
+    int num_meshes;
+    int verts_per_mesh;
+    int num_nodes;
+    int num_materials;
+    int num_animations;
+    int keyframes;
+};
+
+/* ------------------------------------------------------------------ */
+/* Generate the scene                                                 */
+/* ------------------------------------------------------------------ */
+
+struct AccessorInfo {
+    int buffer_view;
+    int component_type;
+    int count;
+    const char *type;
+    float min_vals[3];
+    float max_vals[3];
+    int min_max_components; /* 0 = none, 1 = scalar, 3 = vec3 */
+};
+
+static void generate_scene(const SceneConfig &cfg,
+                            std::string &out_json,
+                            std::vector<uint8_t> &out_bin) {
+    BinBuffer bin;
+
+    /* Pre-compute sizes */
+    int tris_per_mesh = cfg.verts_per_mesh / 3;
+    if (tris_per_mesh < 1) tris_per_mesh = 1;
+    int actual_verts = tris_per_mesh * 3;
+
+    /*
+     * For each mesh:
+     *   - positions: actual_verts * 3 floats
+     *   - normals:   actual_verts * 3 floats
+     *   - indices:   tris_per_mesh * 3 uint16 (or uint32 if >65535)
+     *
+     * For each animation:
+     *   - time keys:  keyframes floats
+     *   - translation values: keyframes * 3 floats
+     */
+
+    /* Track buffer views and accessors */
+    std::vector<size_t> bv_offsets;
+    std::vector<size_t> bv_lengths;
+    std::vector<AccessorInfo> accessors;
+    int bv_idx = 0;
+
+    bool use_u32_indices = (actual_verts > 65535);
+
+    /* === Mesh data === */
+    for (int m = 0; m < cfg.num_meshes; ++m) {
+        float mesh_offset_x = (float)m * 5.0f;
+
+        /* Positions */
+        size_t pos_off = bin.offset();
+        float pmin[3] = {1e30f, 1e30f, 1e30f};
+        float pmax[3] = {-1e30f, -1e30f, -1e30f};
+        for (int v = 0; v < actual_verts; ++v) {
+            float angle = (float)v / (float)actual_verts * 6.2831853f;
+            float r = 1.0f + 0.3f * sinf(angle * 5.0f);
+            float x = mesh_offset_x + r * cosf(angle);
+            float y = r * sinf(angle);
+            float z = 0.5f * sinf(angle * 3.0f + (float)m);
+            bin.push_float(x); bin.push_float(y); bin.push_float(z);
+            if (x < pmin[0]) pmin[0] = x;
+            if (x > pmax[0]) pmax[0] = x;
+            if (y < pmin[1]) pmin[1] = y;
+            if (y > pmax[1]) pmax[1] = y;
+            if (z < pmin[2]) pmin[2] = z;
+            if (z > pmax[2]) pmax[2] = z;
+        }
+        size_t pos_len = bin.offset() - pos_off;
+        bin.align4();
+        bv_offsets.push_back(pos_off); bv_lengths.push_back(pos_len);
+        int pos_bv = bv_idx++;
+
+        AccessorInfo pos_acc;
+        pos_acc.buffer_view = pos_bv;
+        pos_acc.component_type = 5126; /* FLOAT */
+        pos_acc.count = actual_verts;
+        pos_acc.type = "VEC3";
+        memcpy(pos_acc.min_vals, pmin, sizeof(pmin));
+        memcpy(pos_acc.max_vals, pmax, sizeof(pmax));
+        pos_acc.min_max_components = 3;
+        accessors.push_back(pos_acc);
+
+        /* Normals */
+        size_t norm_off = bin.offset();
+        for (int v = 0; v < actual_verts; ++v) {
+            float angle = (float)v / (float)actual_verts * 6.2831853f;
+            float nx = cosf(angle), ny = sinf(angle), nz = 0.0f;
+            float len = sqrtf(nx*nx + ny*ny + nz*nz);
+            if (len > 0) { nx /= len; ny /= len; nz /= len; }
+            bin.push_float(nx); bin.push_float(ny); bin.push_float(nz);
+        }
+        size_t norm_len = bin.offset() - norm_off;
+        bin.align4();
+        bv_offsets.push_back(norm_off); bv_lengths.push_back(norm_len);
+        int norm_bv = bv_idx++;
+
+        AccessorInfo norm_acc;
+        norm_acc.buffer_view = norm_bv;
+        norm_acc.component_type = 5126;
+        norm_acc.count = actual_verts;
+        norm_acc.type = "VEC3";
+        norm_acc.min_max_components = 0;
+        accessors.push_back(norm_acc);
+
+        /* Indices */
+        size_t idx_off = bin.offset();
+        for (int t = 0; t < tris_per_mesh; ++t) {
+            if (use_u32_indices) {
+                bin.push_u32((uint32_t)(t * 3));
+                bin.push_u32((uint32_t)(t * 3 + 1));
+                bin.push_u32((uint32_t)(t * 3 + 2));
+            } else {
+                bin.push_u16((uint16_t)(t * 3));
+                bin.push_u16((uint16_t)(t * 3 + 1));
+                bin.push_u16((uint16_t)(t * 3 + 2));
+            }
+        }
+        size_t idx_len = bin.offset() - idx_off;
+        bin.align4();
+        bv_offsets.push_back(idx_off); bv_lengths.push_back(idx_len);
+        int idx_bv = bv_idx++;
+
+        AccessorInfo idx_acc;
+        idx_acc.buffer_view = idx_bv;
+        idx_acc.component_type = use_u32_indices ? 5125 : 5123; /* UINT or USHORT */
+        idx_acc.count = tris_per_mesh * 3;
+        idx_acc.type = "SCALAR";
+        idx_acc.min_max_components = 0;
+        accessors.push_back(idx_acc);
+    }
+
+    /* === Animation data === */
+    int anim_time_accessor_start = (int)accessors.size();
+    for (int a = 0; a < cfg.num_animations; ++a) {
+        /* Time keys */
+        size_t time_off = bin.offset();
+        float tmin = 0.0f, tmax = 0.0f;
+        for (int k = 0; k < cfg.keyframes; ++k) {
+            float t = (float)k / (float)(cfg.keyframes - 1) * 10.0f;
+            bin.push_float(t);
+            if (k == 0) tmin = t;
+            tmax = t;
+        }
+        size_t time_len = bin.offset() - time_off;
+        bin.align4();
+        bv_offsets.push_back(time_off); bv_lengths.push_back(time_len);
+        int time_bv = bv_idx++;
+
+        AccessorInfo time_acc;
+        time_acc.buffer_view = time_bv;
+        time_acc.component_type = 5126;
+        time_acc.count = cfg.keyframes;
+        time_acc.type = "SCALAR";
+        time_acc.min_vals[0] = tmin;
+        time_acc.max_vals[0] = tmax;
+        time_acc.min_max_components = 1;
+        accessors.push_back(time_acc);
+
+        /* Translation values */
+        size_t val_off = bin.offset();
+        for (int k = 0; k < cfg.keyframes; ++k) {
+            float t = (float)k / (float)(cfg.keyframes - 1) * 10.0f;
+            float x = sinf(t * 0.5f + (float)a) * 2.0f;
+            float y = cosf(t * 0.3f) * 1.5f;
+            float z = sinf(t * 0.7f + (float)a * 0.5f);
+            bin.push_float(x); bin.push_float(y); bin.push_float(z);
+        }
+        size_t val_len = bin.offset() - val_off;
+        bin.align4();
+        bv_offsets.push_back(val_off); bv_lengths.push_back(val_len);
+        int val_bv = bv_idx++;
+
+        AccessorInfo val_acc;
+        val_acc.buffer_view = val_bv;
+        val_acc.component_type = 5126;
+        val_acc.count = cfg.keyframes;
+        val_acc.type = "VEC3";
+        val_acc.min_max_components = 0;
+        accessors.push_back(val_acc);
+    }
+
+    size_t total_bin = bin.data.size();
+
+    /* === Build JSON === */
+    JsonWriter w;
+    w.begin_obj();
+
+    /* asset */
+    w.key("asset"); w.begin_obj();
+    w.kv_str("version", "2.0");
+    w.kv_str("generator", "tinygltf_benchmark_gen");
+    w.end_obj();
+
+    /* scene */
+    w.kv_int("scene", 0);
+
+    /* scenes */
+    w.key("scenes"); w.begin_arr();
+    w.begin_obj();
+    w.kv_str("name", "BenchmarkScene");
+    w.key("nodes"); w.begin_arr();
+    for (int n = 0; n < cfg.num_nodes; ++n) w.val_int(n);
+    w.end_arr();
+    w.end_obj();
+    w.end_arr();
+
+    /* nodes */
+    w.key("nodes"); w.begin_arr();
+    for (int n = 0; n < cfg.num_nodes; ++n) {
+        w.begin_obj();
+        w.kv_str("name", ("Node_" + std::to_string(n)).c_str());
+        if (n < cfg.num_meshes) {
+            w.kv_int("mesh", n);
+        }
+        w.key("translation"); w.begin_arr();
+        w.val_double((double)n * 3.0);
+        w.val_double(0.0);
+        w.val_double(0.0);
+        w.end_arr();
+        w.end_obj();
+    }
+    w.end_arr();
+
+    /* meshes */
+    w.key("meshes"); w.begin_arr();
+    for (int m = 0; m < cfg.num_meshes; ++m) {
+        int base_acc = m * 3; /* pos, norm, idx per mesh */
+        w.begin_obj();
+        w.kv_str("name", ("Mesh_" + std::to_string(m)).c_str());
+        w.key("primitives"); w.begin_arr();
+        w.begin_obj();
+        w.key("attributes"); w.begin_obj();
+        w.kv_int("POSITION", base_acc);
+        w.kv_int("NORMAL", base_acc + 1);
+        w.end_obj();
+        w.kv_int("indices", base_acc + 2);
+        w.kv_int("material", m % cfg.num_materials);
+        w.kv_int("mode", 4);
+        w.end_obj();
+        w.end_arr();
+        w.end_obj();
+    }
+    w.end_arr();
+
+    /* materials */
+    w.key("materials"); w.begin_arr();
+    for (int m = 0; m < cfg.num_materials; ++m) {
+        w.begin_obj();
+        w.kv_str("name", ("Material_" + std::to_string(m)).c_str());
+        w.key("pbrMetallicRoughness"); w.begin_obj();
+        w.key("baseColorFactor"); w.begin_arr();
+        float hue = (float)m / (float)cfg.num_materials;
+        w.val_double(0.5 + 0.5 * sin(hue * 6.28));
+        w.val_double(0.5 + 0.5 * sin(hue * 6.28 + 2.09));
+        w.val_double(0.5 + 0.5 * sin(hue * 6.28 + 4.19));
+        w.val_double(1.0);
+        w.end_arr();
+        w.kv_double("metallicFactor", 0.2 + 0.6 * ((double)m / cfg.num_materials));
+        w.kv_double("roughnessFactor", 0.3 + 0.5 * ((double)(cfg.num_materials - m) / cfg.num_materials));
+        w.end_obj();
+        w.end_obj();
+    }
+    w.end_arr();
+
+    /* accessors */
+    w.key("accessors"); w.begin_arr();
+    for (size_t i = 0; i < accessors.size(); ++i) {
+        const AccessorInfo &a = accessors[i];
+        w.begin_obj();
+        w.kv_int("bufferView", a.buffer_view);
+        w.kv_int("componentType", a.component_type);
+        w.kv_int("count", a.count);
+        w.kv_str("type", a.type);
+        if (a.min_max_components == 1) {
+            w.key("min"); w.begin_arr(); w.val_double(a.min_vals[0]); w.end_arr();
+            w.key("max"); w.begin_arr(); w.val_double(a.max_vals[0]); w.end_arr();
+        } else if (a.min_max_components == 3) {
+            w.key("min"); w.begin_arr();
+            w.val_double(a.min_vals[0]); w.val_double(a.min_vals[1]); w.val_double(a.min_vals[2]);
+            w.end_arr();
+            w.key("max"); w.begin_arr();
+            w.val_double(a.max_vals[0]); w.val_double(a.max_vals[1]); w.val_double(a.max_vals[2]);
+            w.end_arr();
+        }
+        w.end_obj();
+    }
+    w.end_arr();
+
+    /* bufferViews */
+    w.key("bufferViews"); w.begin_arr();
+    for (int i = 0; i < bv_idx; ++i) {
+        w.begin_obj();
+        w.kv_int("buffer", 0);
+        w.kv_int("byteOffset", (int64_t)bv_offsets[i]);
+        w.kv_int("byteLength", (int64_t)bv_lengths[i]);
+        w.end_obj();
+    }
+    w.end_arr();
+
+    /* buffers */
+    w.key("buffers"); w.begin_arr();
+    w.begin_obj();
+    w.kv_int("byteLength", (int64_t)total_bin);
+    /* URI will be set by caller for .gltf, omitted for .glb */
+    w.end_obj();
+    w.end_arr();
+
+    /* animations */
+    if (cfg.num_animations > 0) {
+        w.key("animations"); w.begin_arr();
+        for (int a = 0; a < cfg.num_animations; ++a) {
+            int time_acc = anim_time_accessor_start + a * 2;
+            int val_acc = time_acc + 1;
+            /* Target node: cycle through available nodes */
+            int target_node = a % cfg.num_nodes;
+
+            w.begin_obj();
+            w.kv_str("name", ("Anim_" + std::to_string(a)).c_str());
+
+            w.key("channels"); w.begin_arr();
+            w.begin_obj();
+            w.kv_int("sampler", 0);
+            w.key("target"); w.begin_obj();
+            w.kv_int("node", target_node);
+            w.kv_str("path", "translation");
+            w.end_obj();
+            w.end_obj();
+            w.end_arr();
+
+            w.key("samplers"); w.begin_arr();
+            w.begin_obj();
+            w.kv_int("input", time_acc);
+            w.kv_int("output", val_acc);
+            w.kv_str("interpolation", "LINEAR");
+            w.end_obj();
+            w.end_arr();
+
+            w.end_obj();
+        }
+        w.end_arr();
+    }
+
+    w.end_obj();
+
+    out_json = w.buf;
+    out_bin = bin.data;
+}
+
+/* ------------------------------------------------------------------ */
+/* Write .gltf + .bin                                                 */
+/* ------------------------------------------------------------------ */
+
+static void write_gltf(const std::string &prefix, const std::string &label,
+                        const std::string &json_str,
+                        const std::vector<uint8_t> &bin_data) {
+    std::string bin_name = prefix + "_" + label + ".bin";
+    std::string gltf_name = prefix + "_" + label + ".gltf";
+
+    /* Inject "uri" into the buffer object in JSON */
+    std::string json_patched = json_str;
+    /* Find the buffers array and add uri before the closing } of the buffer */
+    size_t pos = json_patched.find("\"byteLength\"");
+    if (pos != std::string::npos) {
+        /* Find the line end after byteLength value */
+        size_t line_end = json_patched.find('\n', pos);
+        if (line_end != std::string::npos) {
+            /* Extract just the filename for uri */
+            std::string bin_filename = prefix + "_" + label + ".bin";
+            std::string uri_line = ",\n      \"uri\": \"" + bin_filename + "\"";
+            json_patched.insert(line_end, uri_line);
+        }
+    }
+
+    /* Write .bin */
+    FILE *f = fopen(bin_name.c_str(), "wb");
+    if (f) {
+        fwrite(bin_data.data(), 1, bin_data.size(), f);
+        fclose(f);
+    }
+
+    /* Write .gltf */
+    f = fopen(gltf_name.c_str(), "w");
+    if (f) {
+        fwrite(json_patched.c_str(), 1, json_patched.size(), f);
+        fclose(f);
+    }
+
+    printf("  Written: %s (%zu bytes JSON) + %s (%zu bytes binary)\n",
+           gltf_name.c_str(), json_patched.size(),
+           bin_name.c_str(), bin_data.size());
+}
+
+/* ------------------------------------------------------------------ */
+/* Write .glb                                                         */
+/* ------------------------------------------------------------------ */
+
+static void write_glb(const std::string &prefix, const std::string &label,
+                       const std::string &json_str,
+                       const std::vector<uint8_t> &bin_data) {
+    std::string glb_name = prefix + "_" + label + ".glb";
+
+    uint32_t json_len = (uint32_t)json_str.size();
+    uint32_t json_padded = (json_len + 3) & ~3u;
+    uint32_t bin_len = (uint32_t)bin_data.size();
+    uint32_t bin_padded = (bin_len + 3) & ~3u;
+
+    uint32_t total = 12 + 8 + json_padded + 8 + bin_padded;
+
+    FILE *f = fopen(glb_name.c_str(), "wb");
+    if (!f) return;
+
+    /* Header */
+    fwrite("glTF", 1, 4, f);
+    uint32_t version = 2;
+    fwrite(&version, 4, 1, f);
+    fwrite(&total, 4, 1, f);
+
+    /* JSON chunk */
+    uint32_t json_type = 0x4E4F534A;
+    fwrite(&json_padded, 4, 1, f);
+    fwrite(&json_type, 4, 1, f);
+    fwrite(json_str.c_str(), 1, json_len, f);
+    for (uint32_t i = json_len; i < json_padded; ++i) {
+        char sp = ' ';
+        fwrite(&sp, 1, 1, f);
+    }
+
+    /* BIN chunk */
+    uint32_t bin_type = 0x004E4942;
+    fwrite(&bin_padded, 4, 1, f);
+    fwrite(&bin_type, 4, 1, f);
+    fwrite(bin_data.data(), 1, bin_len, f);
+    for (uint32_t i = bin_len; i < bin_padded; ++i) {
+        char z = 0;
+        fwrite(&z, 1, 1, f);
+    }
+
+    fclose(f);
+    printf("  Written: %s (%u bytes)\n", glb_name.c_str(), total);
+}
+
+/* ------------------------------------------------------------------ */
+/* Preset configurations                                              */
+/* ------------------------------------------------------------------ */
+
+struct Preset {
+    const char *label;
+    SceneConfig cfg;
+};
+
+static Preset presets[] = {
+    {"tiny",   {1,   100,    2,  1, 0,   0}},
+    {"small",  {5,   1000,  10,  3, 2,  50}},
+    {"medium", {20,  5000,  50, 10, 5, 200}},
+    {"large",  {100, 10000, 200, 20, 10, 500}},
+    {"huge",   {500, 50000, 1000, 50, 50, 1000}},
+};
+
+static const int num_presets = (int)(sizeof(presets) / sizeof(presets[0]));
+
+/* ------------------------------------------------------------------ */
+/* Generate float-heavy scene (~500MB of ASCII float values in JSON)  */
+/* ------------------------------------------------------------------ */
+
+static void generate_float_heavy(const std::string &prefix, size_t target_mb) {
+    std::string gltf_name = prefix + "_float_heavy.gltf";
+    FILE *f = fopen(gltf_name.c_str(), "w");
+    if (!f) {
+        fprintf(stderr, "Cannot open %s\n", gltf_name.c_str());
+        return;
+    }
+
+    /* Write minimal valid glTF with massive extras float array */
+    fprintf(f, "{\n");
+    fprintf(f, "  \"asset\": {\"version\": \"2.0\", \"generator\": \"tinygltf_benchmark_gen\"},\n");
+    fprintf(f, "  \"scene\": 0,\n");
+    fprintf(f, "  \"scenes\": [{\"name\": \"FloatHeavy\", \"nodes\": [0]}],\n");
+    fprintf(f, "  \"nodes\": [{\"name\": \"Root\"}],\n");
+    fprintf(f, "  \"extras\": {\n");
+
+    size_t target_bytes = target_mb * 1024ULL * 1024ULL;
+    size_t total_written = 0;
+    int num_channels = 10;
+    size_t per_channel = target_bytes / (size_t)num_channels;
+
+    for (int ch = 0; ch < num_channels; ++ch) {
+        fprintf(f, "    \"channel_%d\": [\n      ", ch);
+        size_t ch_written = 0;
+        size_t count = 0;
+        uint64_t seed = (uint64_t)ch * 7919ULL + 1;
+        bool first = true;
+
+        while (ch_written < per_channel) {
+            /* Comma before every value except the first */
+            if (!first) {
+                fwrite(",\n      ", 1, 8, f);
+                ch_written += 8;
+            }
+            first = false;
+
+            /* Generate varied float values: mix of magnitudes and precisions */
+            seed = seed * 6364136223846793005ULL + 1442695040888963407ULL;
+            double raw = (double)(int64_t)seed / (double)INT64_MAX;
+
+            double val;
+            int kind = (int)(count % 5);
+            switch (kind) {
+                case 0: val = raw * 1000.0; break;           /* large: -999.xxx */
+                case 1: val = raw * 0.001; break;            /* small: 0.000xxx */
+                case 2: val = raw * 3.14159265358979; break; /* medium: -3.14..3.14 */
+                case 3: val = raw * 1e6; break;              /* very large */
+                case 4: val = raw * 1e-6; break;             /* very small */
+                default: val = raw; break;
+            }
+
+            char buf[64];
+            int len = snprintf(buf, sizeof(buf), "%.8g", val);
+            fwrite(buf, 1, (size_t)len, f);
+            ch_written += (size_t)len;
+            count++;
+        }
+
+        total_written += ch_written;
+
+        if (ch < num_channels - 1) {
+            fprintf(f, "\n    ],\n");
+        } else {
+            fprintf(f, "\n    ]\n");
+        }
+    }
+
+    fprintf(f, "  }\n");
+    fprintf(f, "}\n");
+    fclose(f);
+
+    /* Report actual file size */
+    f = fopen(gltf_name.c_str(), "rb");
+    if (f) {
+        fseek(f, 0, SEEK_END);
+        long sz = ftell(f);
+        fclose(f);
+        printf("  Written: %s (%.1f MB, ~%zu float values across %d channels)\n",
+               gltf_name.c_str(), (double)sz / (1024.0 * 1024.0),
+               total_written / 12, num_channels);
+    }
+}
+
+/* ------------------------------------------------------------------ */
+/* Main                                                               */
+/* ------------------------------------------------------------------ */
+
+int main(int argc, char **argv) {
+    std::string prefix = "synthetic";
+
+    /* Parse args */
+    for (int i = 1; i < argc; ++i) {
+        if (strcmp(argv[i], "--prefix") == 0 && i + 1 < argc) {
+            prefix = argv[++i];
+        }
+    }
+
+    printf("Generating synthetic glTF benchmark scenes...\n\n");
+
+    for (int p = 0; p < num_presets; ++p) {
+        const Preset &pr = presets[p];
+
+        printf("[%s] meshes=%d verts/mesh=%d nodes=%d materials=%d "
+               "animations=%d keyframes=%d\n",
+               pr.label, pr.cfg.num_meshes, pr.cfg.verts_per_mesh,
+               pr.cfg.num_nodes, pr.cfg.num_materials,
+               pr.cfg.num_animations, pr.cfg.keyframes);
+
+        std::string json;
+        std::vector<uint8_t> bin;
+        generate_scene(pr.cfg, json, bin);
+
+        write_gltf(prefix, pr.label, json, bin);
+        write_glb(prefix, pr.label, json, bin);
+        printf("\n");
+    }
+
+    /* Float-heavy scene: ~500MB of ASCII floats in JSON */
+    printf("[float_heavy] ~500MB of ASCII float values in JSON extras\n");
+    generate_float_heavy(prefix, 500);
+    printf("\n");
+
+    printf("Done.\n");
+    return 0;
+}

tests/tester.cc

@@ -1121,8 +1121,10 @@ TEST_CASE("images-as-is", "[issue-487]") {
     // All the images should have been written to disk with their original data
     for (const auto& image : model.images)  {
       // Make sure the image files exist
-      std::fstream file(image.uri);
-      CHECK(file.good());
+      {
+        std::fstream file(image.uri);
+        CHECK(file.good());
+      } // Close file before stbi_load (Windows sharing violation fix)
 #ifndef TINYGLTF_NO_STB_IMAGE
       // Make sure we can load the images
       int w = -1, h = -1, component = -1;

tests/v3/fuzzer/Makefile

@@ -0,0 +1,67 @@
+# tests/v3/fuzzer/Makefile — Build libFuzzer harness for tinygltf v3
+#
+# Requires: clang++ with libFuzzer support
+#
+# Targets:
+#   make             — build fuzzer with ASan + UBSan
+#   make run         — run fuzzer with default settings
+#   make seed        — generate seed corpus from test models
+#   make clean       — remove binaries and corpus
+
+CXX       = clang++
+CXXFLAGS  = -g -O1 -std=c++17 -fno-rtti -fno-exceptions
+SANITIZE  = -fsanitize=fuzzer,address,undefined
+INCLUDES  = -I../../..
+
+FUZZER    = fuzz_gltf_v3
+CORPUS    = corpus
+ARTIFACTS = artifacts
+
+# Fuzzer runtime options
+MAX_LEN   ?= 65536
+JOBS      ?= $(shell nproc 2>/dev/null || echo 4)
+MAX_TIME  ?= 0
+
+.PHONY: all run seed clean
+
+all: $(FUZZER)
+
+$(FUZZER): fuzz_gltf_v3.cc ../../../tiny_gltf_v3.h ../../../tinygltf_json.h
+	$(CXX) $(CXXFLAGS) $(SANITIZE) $(INCLUDES) -o $@ $<
+
+run: $(FUZZER) | $(CORPUS) $(ARTIFACTS)
+	./$(FUZZER) $(CORPUS) \
+		-artifact_prefix=$(ARTIFACTS)/ \
+		-max_len=$(MAX_LEN) \
+		-jobs=$(JOBS) \
+		-workers=$(JOBS) \
+		$(if $(filter-out 0,$(MAX_TIME)),-max_total_time=$(MAX_TIME))
+
+# Generate seed corpus from existing test models
+seed: | $(CORPUS)
+	@echo "Seeding corpus from test models..."
+	@for f in ../../../models/Cube/Cube.gltf \
+	          ../../../models/Cube/Cube.glb; do \
+		if [ -f "$$f" ]; then \
+			cp "$$f" $(CORPUS)/; \
+			echo "  Added: $$f"; \
+		fi; \
+	done
+	@# Add a minimal valid glTF JSON
+	@echo '{"asset":{"version":"2.0"},"scene":0,"scenes":[{"nodes":[0]}],"nodes":[{"name":"n"}]}' > $(CORPUS)/minimal.gltf
+	@# Add a minimal valid GLB (header + empty JSON chunk)
+	@printf 'glTF\x02\x00\x00\x00\x1c\x00\x00\x00\x04\x00\x00\x00JSON{}  ' > $(CORPUS)/minimal.glb
+	@# Add edge cases
+	@echo '{}' > $(CORPUS)/empty_object.gltf
+	@echo '{"asset":{"version":"2.0"}}' > $(CORPUS)/asset_only.gltf
+	@echo "Corpus: $$(ls $(CORPUS) | wc -l) files"
+
+$(CORPUS):
+	mkdir -p $(CORPUS)
+
+$(ARTIFACTS):
+	mkdir -p $(ARTIFACTS)
+
+clean:
+	rm -f $(FUZZER)
+	rm -rf $(CORPUS) $(ARTIFACTS)

tests/v3/fuzzer/fuzz_gltf_v3.cc

@@ -0,0 +1,110 @@
+/*
+ * fuzz_gltf_v3.cc — libFuzzer harness for tinygltf v3 parser.
+ *
+ * Fuzz targets:
+ *   - Auto-detect (GLB or JSON) parse from arbitrary bytes
+ *   - Exercises JSON parser, GLB header parsing, arena allocator,
+ *     error stack, and all glTF entity parsing paths.
+ *
+ * Build (clang with libFuzzer):
+ *   clang++ -g -O1 -fsanitize=fuzzer,address,undefined \
+ *       -std=c++17 -fno-rtti -fno-exceptions \
+ *       -I../../.. -o fuzz_gltf_v3 fuzz_gltf_v3.cc
+ *
+ * Run:
+ *   ./fuzz_gltf_v3 corpus/ -max_len=65536
+ *
+ * Seed corpus: place valid .gltf and .glb files in corpus/
+ */
+
+#define TINYGLTF3_IMPLEMENTATION
+#include "tiny_gltf_v3.h"
+
+#include <cstdint>
+#include <cstddef>
+
+/* Memory budget to prevent OOM during fuzzing */
+static const uint64_t FUZZ_MEMORY_BUDGET = 64ULL * 1024 * 1024; /* 64 MB */
+
+static void fuzz_parse_auto(const uint8_t *data, size_t size) {
+    tg3_model model;
+    tg3_error_stack errors;
+    tg3_error_stack_init(&errors);
+
+    tg3_parse_options opts;
+    tg3_parse_options_init(&opts);
+    opts.memory.memory_budget = FUZZ_MEMORY_BUDGET;
+
+    tg3_parse_auto(&model, &errors, data, (uint64_t)size,
+                   "", 0, &opts);
+
+    tg3_model_free(&model);
+    tg3_error_stack_free(&errors);
+}
+
+static void fuzz_parse_json(const uint8_t *data, size_t size) {
+    tg3_model model;
+    tg3_error_stack errors;
+    tg3_error_stack_init(&errors);
+
+    tg3_parse_options opts;
+    tg3_parse_options_init(&opts);
+    opts.memory.memory_budget = FUZZ_MEMORY_BUDGET;
+
+    tg3_parse(&model, &errors, data, (uint64_t)size,
+              "", 0, &opts);
+
+    tg3_model_free(&model);
+    tg3_error_stack_free(&errors);
+}
+
+static void fuzz_parse_glb(const uint8_t *data, size_t size) {
+    tg3_model model;
+    tg3_error_stack errors;
+    tg3_error_stack_init(&errors);
+
+    tg3_parse_options opts;
+    tg3_parse_options_init(&opts);
+    opts.memory.memory_budget = FUZZ_MEMORY_BUDGET;
+
+    tg3_parse_glb(&model, &errors, data, (uint64_t)size,
+                  "", 0, &opts);
+
+    tg3_model_free(&model);
+    tg3_error_stack_free(&errors);
+}
+
+static void fuzz_parse_float32(const uint8_t *data, size_t size) {
+    tg3_model model;
+    tg3_error_stack errors;
+    tg3_error_stack_init(&errors);
+
+    tg3_parse_options opts;
+    tg3_parse_options_init(&opts);
+    opts.memory.memory_budget = FUZZ_MEMORY_BUDGET;
+    opts.parse_float32 = 1;
+
+    tg3_parse_auto(&model, &errors, data, (uint64_t)size,
+                   "", 0, &opts);
+
+    tg3_model_free(&model);
+    tg3_error_stack_free(&errors);
+}
+
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+    if (size == 0) return 0;
+
+    /* Use first byte to select parse path, rest is the payload */
+    uint8_t selector = data[0] % 4;
+    const uint8_t *payload = data + 1;
+    size_t payload_size = size - 1;
+
+    switch (selector) {
+    case 0: fuzz_parse_auto(payload, payload_size);    break;
+    case 1: fuzz_parse_json(payload, payload_size);    break;
+    case 2: fuzz_parse_glb(payload, payload_size);     break;
+    case 3: fuzz_parse_float32(payload, payload_size); break;
+    }
+
+    return 0;
+}

tiny_gltf.h

@@ -50,12 +50,6 @@
 #include <utility>
 #include <vector>
 
-// Auto-detect C++14 standard version
-#if !defined(TINYGLTF_USE_CPP14) && defined(__cplusplus) && \
-    (__cplusplus >= 201402L)
-#define TINYGLTF_USE_CPP14
-#endif
-
 #ifdef __ANDROID__
 #ifdef TINYGLTF_ANDROID_LOAD_FROM_ASSETS
 #include <android/asset_manager.h>
@@ -834,20 +828,20 @@ struct Accessor {
       maxValues;  // optional. integer value is promoted to double
 
   struct Sparse {
-    int count;
-    bool isSparse;
+    int count{0};
+    bool isSparse{false};
     struct {
-      size_t byteOffset;
-      int bufferView;
-      int componentType;  // a TINYGLTF_COMPONENT_TYPE_ value
+      size_t byteOffset{0};
+      int bufferView{-1};
+      int componentType{-1};  // a TINYGLTF_COMPONENT_TYPE_ value
       Value extras;
       ExtensionMap extensions;
       std::string extras_json_string;
       std::string extensions_json_string;
     } indices;
     struct {
-      int bufferView;
-      size_t byteOffset;
+      int bufferView{-1};
+      size_t byteOffset{0};
       Value extras;
       ExtensionMap extensions;
       std::string extras_json_string;
@@ -898,11 +892,7 @@ struct Accessor {
     // unreachable return 0;
   }
 
-  Accessor()
-
-  {
-    sparse.isSparse = false;
-  }
+  Accessor() = default;
   DEFAULT_METHODS(Accessor)
   bool operator==(const tinygltf::Accessor &) const;
 };
@@ -1721,7 +1711,11 @@ class TinyGLTF {
 #endif  // __GNUC__
 
 #ifndef TINYGLTF_NO_INCLUDE_JSON
-#ifndef TINYGLTF_USE_RAPIDJSON
+#ifdef TINYGLTF_USE_CUSTOM_JSON
+#ifndef TINYGLTF_NO_INCLUDE_CUSTOM_JSON
+#include "tinygltf_json.h"
+#endif
+#elif !defined(TINYGLTF_USE_RAPIDJSON)
 #include "json.hpp"
 #else
 #ifndef TINYGLTF_NO_INCLUDE_RAPIDJSON
@@ -1809,7 +1803,10 @@ class TinyGLTF {
 
 namespace tinygltf {
 namespace detail {
-#ifdef TINYGLTF_USE_RAPIDJSON
+#ifdef TINYGLTF_USE_CUSTOM_JSON
+// Types and JsonParse are provided by tinygltf_json.h (already included above)
+// via 'namespace tinygltf { namespace detail { ... } }' declarations.
+#elif defined(TINYGLTF_USE_RAPIDJSON)
 
 #ifdef TINYGLTF_USE_RAPIDJSON_CRTALLOCATOR
 // This uses the RapidJSON CRTAllocator.  It is thread safe and multiple
@@ -1881,6 +1878,7 @@ using json_const_array_iterator = json_const_iterator;
 using JsonDocument = json;
 #endif
 
+#ifndef TINYGLTF_USE_CUSTOM_JSON
 void JsonParse(JsonDocument &doc, const char *str, size_t length,
                bool throwExc = false) {
 #ifdef TINYGLTF_USE_RAPIDJSON
@@ -1890,6 +1888,7 @@ void JsonParse(JsonDocument &doc, const char *str, size_t length,
   doc = detail::json::parse(str, str + length, nullptr, throwExc);
 #endif
 }
+#endif  // !TINYGLTF_USE_CUSTOM_JSON
 }  // namespace detail
 }  // namespace tinygltf
 
@@ -2461,9 +2460,8 @@ inline unsigned char from_hex(unsigned char ch) {
 }
 
 static const std::string urldecode(const std::string &str) {
-  using namespace std;
-  string result;
-  string::size_type i;
+  std::string result;
+  std::string::size_type i;
   for (i = 0; i < str.size(); ++i) {
     if (str[i] == '+') {
       result += ' ';
@@ -3447,6 +3445,7 @@ bool DecodeDataURI(std::vector<unsigned char> *out, std::string &mime_type,
   return true;
 }
 
+#ifndef TINYGLTF_USE_CUSTOM_JSON
 namespace detail {
 bool GetInt(const detail::json &o, int &val) {
 #ifdef TINYGLTF_USE_RAPIDJSON
@@ -3670,6 +3669,7 @@ std::string JsonToString(const detail::json &o, int spacing = -1) {
 }
 
 }  // namespace detail
+#endif  // !TINYGLTF_USE_CUSTOM_JSON
 
 static bool ParseJsonAsValue(Value *ret, const detail::json &o) {
   Value val{};
@@ -4362,7 +4362,7 @@ static bool ParseImage(Image *image, const int image_idx, std::string *err,
     // Just only save some information here. Loading actual image data from
     // bufferView is done after this `ParseImage` function.
     image->bufferView = bufferView;
-    image->mimeType = mime_type;
+    image->mimeType = std::move( mime_type );
     image->width = width;
     image->height = height;
 
@@ -5261,7 +5261,7 @@ static bool ParseNode(Node *node, std::string *err, const detail::json &o,
   if (node->extensions.count("MSFT_lod") != 0) {
     auto const &msft_lod_ext = node->extensions["MSFT_lod"];
     if (msft_lod_ext.Has("ids")) {
-      auto idsArr = msft_lod_ext.Get("ids");
+      const auto &idsArr = msft_lod_ext.Get("ids");
       for (size_t i = 0; i < idsArr.ArrayLen(); ++i) {
         node->lods.emplace_back(idsArr.Get(i).GetNumberAsInt());
       }
@@ -5290,7 +5290,7 @@ static bool ParseScene(Scene *scene, std::string *err, const detail::json &o,
   if (scene->extensions.count("KHR_audio") != 0) {
     auto const &audio_ext = scene->extensions["KHR_audio"];
     if (audio_ext.Has("emitters")) {
-      auto emittersArr = audio_ext.Get("emitters");
+      const auto &emittersArr = audio_ext.Get("emitters");
       for (size_t i = 0; i < emittersArr.ArrayLen(); ++i) {
         scene->audioEmitters.emplace_back(emittersArr.Get(i).GetNumberAsInt());
       }
@@ -5326,7 +5326,7 @@ static bool ParsePbrMetallicRoughness(
       }
       return false;
     }
-    pbr->baseColorFactor = baseColorFactor;
+    pbr->baseColorFactor = std::move( baseColorFactor );
   }
 
   {
@@ -5478,7 +5478,7 @@ static bool ParseMaterial(Material *material, std::string *err, std::string *war
   if (material->extensions.count("MSFT_lod") != 0) {
     auto const &msft_lod_ext = material->extensions["MSFT_lod"];
     if (msft_lod_ext.Has("ids")) {
-      auto idsArr = msft_lod_ext.Get("ids");
+      const auto &idsArr = msft_lod_ext.Get("ids");
       for (size_t i = 0; i < idsArr.ArrayLen(); ++i) {
         material->lods.emplace_back(idsArr.Get(i).GetNumberAsInt());
       }
@@ -6453,6 +6453,15 @@ bool TinyGLTF::LoadFromString(Model *model, std::string *err, std::string *warn,
           return false;
         }
         const Buffer &buffer = model->buffers[size_t(bufferView.buffer)];
+        if (bufferView.byteOffset >= buffer.data.size()) {
+          if (err) {
+            std::stringstream ss;
+            ss << "image[" << idx << "] bufferView \"" << image.bufferView
+               << "\" indexed out of bounds of its buffer." << std::endl;
+            (*err) += ss.str();
+          }
+          return false;
+        }
 
         if (LoadImageData == nullptr) {
           if (err) {
@@ -6964,6 +6973,7 @@ bool TinyGLTF::LoadBinaryFromFile(Model *model, std::string *err,
 ///////////////////////
 // GLTF Serialization
 ///////////////////////
+#ifndef TINYGLTF_USE_CUSTOM_JSON
 namespace detail {
 detail::json JsonFromString(const char *s) {
 #ifdef TINYGLTF_USE_RAPIDJSON
@@ -7036,6 +7046,7 @@ void JsonReserveArray(detail::json &o, size_t s) {
   (void)(s);
 }
 }  // namespace detail
+#endif  // !TINYGLTF_USE_CUSTOM_JSON
 
 // typedef std::pair<std::string, detail::json> json_object_pair;