This project reimplements inference components described by and released with
FlashVSR: Towards Real-Time Diffusion-Based Streaming Video Super-Resolution.

Original project: https://github.com/OpenImagingLab/FlashVSR
ComfyUI reference: https://github.com/Comfy-Org/ComfyUI
Distributed safetensors: https://huggingface.co/pizzawookiee/FlashVSR-1.1

This project is distributed under GPL-3.0-only. During development, behavior
and integration approaches were compared with the following GPL-3.0 projects:
https://github.com/Comfy-Org/ComfyUI
https://github.com/1038lab/ComfyUI-FlashVSR
https://github.com/lihaoyun6/ComfyUI-FlashVSR_Ultra_Fast
https://github.com/kijai/ComfyUI-KJNodes

GPL-3.0-only is used as the conservative project-wide license. It is compatible
with incorporating or adapting Apache-2.0 material under GPLv3's terms; this
notice does not relicense any third-party work.

WanVideoWrapper reference: https://github.com/kijai/ComfyUI-WanVideoWrapper
WanVideoWrapper is distributed under the Apache License 2.0.

The optional block-sparse attention adapter normally calls SpargeAttn's public
block_sparse_sage2_attn_cuda API. When CUDA profiling is enabled, it mirrors
that Apache-2.0 public function's execution sequence and calls the packaged
quantization, LUT, V-layout, and architecture-specific kernel helpers directly
so their costs can be measured separately:
https://github.com/woct0rdho/SpargeAttn
Copyright (c) 2025 Jintao Zhang, Chendong Xiang, Haofeng Huang.
SpargeAttn is distributed separately under the Apache License 2.0.

Version 0.34 additionally calls SpargeAttn's packaged architecture-specific
lower symbols after preparing compatible Q/K/V tensors. The row-INT8 cache
conversion, fixed-reference K preparation, direct transposed-FP8 V
materialization, capability dispatch, and fallback logic in this repository
are original implementations. No SpargeAttn CUDA kernel source is bundled.

The temporary-attention routing design was informed by ComfyUI-RadialAttn:
https://github.com/woct0rdho/ComfyUI-RadialAttn
Copyright (c) 2025, Apache License 2.0.

The feasibility of reusing quantized projection inputs was evaluated against
the documented behavior of H3-Optimizations:
https://github.com/Zironic/H3-Optimizations
No H3-Optimizations source code is incorporated. The FlashVSR QKV and carrier
implementation in this repository was written independently against ComfyUI
and Comfy-Kitchen interfaces.

The optional cache-residency backend uses the public ModelVBAR interfaces
provided with ComfyUI's comfy-aimdo integration:
https://github.com/Comfy-Org/comfy-aimdo
comfy-aimdo is distributed under GPL-3.0. No comfy-aimdo source code is
bundled in this repository.
