Without sparse attention, the model uses full (dense) attention which
attends to distant irrelevant information, causing ghosting artifacts.
The FlashVSR paper explicitly requires block-sparse attention.
Vendored from SageAttention team (Apache 2.0), pure Triton (no CUDA C++).
Import chain: local sparse_sage → external sageattn.core → SDPA fallback.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The local_attn_mask was not being tiled across temporal dimensions,
causing assertion errors in streaming mode and wrong masks otherwise.
Match naxci1 reference: 4D tile/rearrange for Q/K temporal windows,
chunk-based score computation, and topk<=0 guard.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Use generate_draft_block_mask_refined for sparse attention mask (matches
naxci1's generate_draft_block_mask_sage with proper half-block key scoring)
- Remove spurious repeat_interleave(2, dim=-1) from generate_draft_block_mask
that doubled the key dimension incorrectly
- Add torch.clamp(0, 1) to _to_frames output (matches naxci1's tensor2video)
- Add .to(self.device) on LQ video slices in streaming loop for all pipelines
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Root cause of remaining ghosting: our single-stage temporal padding
(N+4 → floor to 8k+1) TRUNCATED frames when N+4 wasn't already 8k+1.
For 50 frames: 50+4=54 → floor to 49, LOSING the last input frame.
The pipeline then processed misaligned LQ→output frame mapping.
Fix matches naxci1/ComfyUI-FlashVSR_Stable two-stage approach:
1. Pad to next_8n5(N) (next integer >= N of form 8k+5, minimum 21)
2. Add 4 → result is always 8(k+1)+1, a valid 8k+1 — NEVER truncates
Also:
- kv_ratio default 2.0→3.0 (matches naxci1, max quality KV cache)
- local_range default 9→11 (more stable temporal consistency)
- sinusoidal_embedding_1d, precompute_freqs_cis, rope_apply: float32→float64
(matches naxci1 reference precision for embeddings and RoPE)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When sageattn fails, q/k/v are already in [b,n,s,d] format from the
rearrange before the call. Use SDPA directly on them instead of calling
_sdpa_fallback which expects [b,s,(n*d)] and crashes with a shape error.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
SageAttention CUDA kernels don't support Blackwell yet. Catch runtime
failures from sageattn/sparse_sageattn, disable them, and fall back to
PyTorch SDPA. Only pays the try/except cost once per session.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Verify attention backend functions are actually callable before marking
them available. Falls back to PyTorch SDPA instead of calling None.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Vendor minimal diffsynth subset for FlashVSR inference (full/tiny pipelines,
v1 and v1.1 checkpoints auto-downloaded from HuggingFace). Includes segment-based
processing with temporal overlap and crossfade blending for bounded RAM on long videos.
Nodes: Load FlashVSR Model, FlashVSR Upscale, FlashVSR Segment Upscale.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
