ComfyUI-FlashVSR Stable version

[naxci1]

### Running FlashVSR on lower VRAM without any artifacts.

### 08.12.2025

Optimization: Ported all VRAM optimizations and Tiled VAE support to tiny-long mode, ensuring feature parity across all modes.
Performance: Optimized for Windows environments (speed improvements).
Update: General codebase cleanup and synchronization.

### 07.12.2025

VRAM Optimization: Implemented auto-fallback for process_chunk. If OOM occurs, it automatically retries with tiled_vae=True and then tiled_dit=True, preventing crashes.
Critical Fix: Fixed a bug in the non-tiled processing path where output was undefined.
Optimization: Defer VAE loading in full mode until strictly necessary, significantly reducing peak VRAM usage.
Optimization: Added a proactive 90% VRAM usage warning.
Refactor: Rewrote progress bar to use single-line in-place updates (\r) for cleaner console output.
Defaults: Updated default settings for FlashVSR Ultra-Fast node to be safer for 16GB cards (unload_dit=True, tiled options enabled).
Bug Fix: Fixed AttributeError in full mode by adding a fallback mechanism to manually load the VAE model if the model manager fails.
Bug Fix: Fixed the progress bar to correctly display status in ComfyUI using the cqdm wrapper. Added text-based progress bar to logs.
Sync: Enabled VAE spatial tiling for tiny mode, bringing VRAM savings from tiny-long to the standard fast pipeline.
Documentation: Expanded tooltips for all node parameters and added detailed usage instructions to README.
New Feature: Added frame_chunk_size option to split large videos into chunks, enabling processing of large files on limited VRAM by offloading to CPU.
Enhancement: Improved logging to show detailed resource usage (RAM, Peak VRAM, per-step timing) and model configuration details.
Optimization: Added torch.cuda.ipc_collect() for better memory cleanup.
New Feature: Added attention_mode selection with support for flash_attention_2, sdpa, sparse_sage, and block_sparse backends.
Refactor: Cleaned up code and improved error handling for imports.

### 06.12.2025

Bug Fix: Fixed a shape mismatch error for small input frames by implementing correct padding logic.
Optimization: VRAM is now immediately freed at the start of processing to prevent OOM errors.
New Feature: Added enable_debug option for extensive logging.
New Feature: Added keep_models_on_cpu option to keep models in RAM (CPU) instead of VRAM.
Enhancement: Added accurate FPS calculation and peak VRAM reporting.
Optimization: Replaced einops operations with native PyTorch ops.
Optimization: Added "Conv3d memory workaround".

https://github.com/naxci1/ComfyUI-FlashVSR_Stable

[naxci1]

[1.2.7] - 2025-12-23

✨ New Features

5 New VAE Models: Added support for Wan2.2, LightVAE_W2.1, TAE_W2.2, LightTAE_HY1.5 (plus original Wan2.1).
Smart Resource Calculator: Added "Pre-flight" check to analyze VRAM/RAM usage and recommend optimal settings before processing.
Auto-Download: Added automatic downloading for missing VAE models from HuggingFace (lightx2v/Autoencoders repository).
Unified UI: Replaced complex vae_type and alt_vae inputs with a single vae_model dropdown.

🐛 Bug Fixes

Fixed Black Borders: Implemented correct "Padding → Process → Crop" logic to handle VAE dimension requirements without corrupting the output.
Fixed Video Glitches: Corrected Tensor Permutation logic (B,C,F,H,W vs B,H,W,C) to prevent visual artifacts.
Fixed Model Loading: Solved the issue where selecting "Wan2.2" incorrectly loaded "Wan2.1_VAE.pth".
Fixed OOM False Positive: OOM recovery no longer triggers prematurely (adjusted threshold to 95%).

⚡ Optimization

Unified Pipeline: Merged tiny, full, and tiny-long logic into a single optimized architecture.
OOM Protection: Improved memory management with 95% VRAM threshold using torch.cuda.mem_get_info().
Lossless Resize: Uses NEAREST interpolation for integer scaling (0.5, 0.25) to avoid blur.
Aggressive Garbage Collection: Added torch.cuda.empty_cache() before/after heavy VAE operations.

🛠 Refactoring

Explicit VAE Instantiation: No more state_dict inspection/guessing - strict class mapping based on user selection.
VAE_MODEL_MAP: Centralized configuration for all 5 VAE options (class, file, URL, dimensions).
Summary Logging: End-of-processing report with total time, peak VRAM, and resolutions.
Debug Logging: Shows selected_model vs loaded_model_path for verification.

📖 Documentation

Updated README with VAE Selection Guide and VRAM tier recommendations.
Added Best Practices section for Low/Medium/High VRAM configurations.
Added Pre-Flight Resource Check documentation with example output.
Added direct download links for all 5 VAE models.