Keep weights and arguments on the same device when 'preserve_vram' is ON
Restored compatibility with FP8 safetensors for MPS backend
Added multi-gpu support for ComfyUI nodes
Added FPS calculation
Set use_non_blocking to False by default to prevent memory leaks (temp fix)
Moved some parms from main node to extra args node
**Architecture & Performance:**
- Unified debug system with categorized logging and memory tracking
- FP8 models now stay in FP8, convert to BF16 only for math operations (faster, less memory)
- Fixed memory leaks
- Removed ComfyUI dependency for standalone compatibility
- Better VRAM management between batches
**New Features:**
- VAE tiling for larger/longer video upscaling
- Multi-repo model support (numz/ and AInVFX/)
- Model auto-discovery in ComfyUI folder
- CLI: BlockSwap options, temporal overlap blending, prepend_frames for better first frames
- `enable_debug` and `cache_model` moved to main node
**Code Organization:**
- New modular structure: `constants.py`, `model_registry.py`, `debug.py`
- Removed legacy code and dead files
- Added mixed FP8 models to fix 7B artifacts
- Fixed issue that would lead to a white border around the image due to
improper tile blending
- Added smoother blending (hann ramp) for the tiles and temporal
overlaps
- Fix memory leak in slicing_decode by cleaning memory of InflatedCausalConv3d
- Prevent OOM errors during final video concatenation by keeping the tensors on CPU
- use h/w tuples to specify tiling size and overlap to allow for finer
control
- added OneOrTwoValues class to handle CLI input of single values or h/w
pairs
- Fix issues with unnecessary VAE blocks at the boundarys
- Updated VAE encode/decode logic to more closely match the original flow
- Better skip_frame log info
- Move cache_model and enable_debug options from BlockSwap to main SeedVR2 node
- cache_model: Keep models in RAM between runs (skip reload for faster iterations)
- enable_debug: Show detailed memory/timing tracking with hierarchical display
- Replace hardcoded debug statements with unified Debug class
- Add visual icons for log categorization
- Implement parent-child timing relationships for operation breakdown
- Add VRAM/RAM usage tracking with before/after comparisons
Model caching now available for all workflows, not just BlockSwap users.
Debug system provides granular performance insights when enabled.
- Add model registry with metadata for all SeedVR2 models in different HF repos
- Support dynamic discovery of .safetensors and .gguf files
- Centralize constants and remove duplicate model lists
- Simplify download logic with clear manual fallback instructions
Adrien Toupet