- Fix AttributeDict handling in TorchCompile node to prevent .copy() errors
- Fix duplicate model loading in GeneratorAdvanced causing memory allocation failures
- Update model loader to properly handle preloaded models from pipeline
- Fix DAC VAE compilation (was looking for 'vae' key instead of 'dac_model')
- Make model_path and config_path optional parameters with proper defaults
- Add better error handling and validation throughout the pipeline
- Prevent unnecessary model reloading when using pre-compiled models
The advanced pipeline now works correctly:
ModelLoader → DependenciesLoader → TorchCompile → GeneratorAdvanced
Co-Authored-By: if@impactframes <if@impactframes.ai>
- Add HunyuanVideoFoleyModelLoader with FP8 weight-only quantization (fp8_e4m3fn, fp8_e5m2)
- Add HunyuanVideoFoleyDependenciesLoader for modular component loading
- Add HunyuanVideoFoleyTorchCompile for ~30% inference speedup
- Add HunyuanVideoFoleyGeneratorAdvanced that accepts pre-loaded models
- Keep original HunyuanVideoFoley node fully functional for backward compatibility
- Fix model path to use ComfyUI's foley/hunyuanvideo-foley-xxl directory structure
- Support both standalone and pipeline-based workflows
Co-Authored-By: if@impactframes <if@impactframes.ai>
- Fix map_location conflict in DAC model loading with fallback manual loading
- Revert _extract_frames_from_image_input() to original PR #3 simple approach
- Fix indentation errors in exception handling blocks
- Resolve black screen issues with VHS node frame input processing
ImpactFrames