Comfyui-Relay-VSR

Comfyui-Relay-VSR

A ComfyUI custom node pack for Relay-VSR and LightX2V Relay-VSR-2x — a one-step distilled, pixel-space video super-resolution model. It runs the Wan backbone and the Flash decoder directly, with no VAE, CLIP or text encoder: images in, images out.

• Native LightX2V 2× weights (RelayVSR_X2)
• Original 4× weights (RelayVSR_X4)
• Streaming windows: only a small number of frames are in flight, so memory stays flat for long clips
• Optional INT8 conv-rotation weights for the Wan backbone
• Optional per-block Wan CPU offload

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/<your-account>/Comfyui-Relay-VSR.git

Restart ComfyUI. No additional pip packages are required: the nodes only use torch, safetensors, numpy and ComfyUI's own dependencies (comfy.quant_ops, comfy-kitchen). There is intentionally no requirements.txt, so ComfyUI's pinned versions stay authoritative.

Models

| Scale | Source | | --- | --- | | 2× | https://huggingface.co/lightx2v/RelayVSR_X2 | | 4× | https://github.com/kopperx/RelayVSR |

Place one directory per bundle under ComfyUI/models/relayvsr/:

ComfyUI/models/relayvsr/
├── RelayVSR_X2/
│   ├── wan_x2.safetensors
│   └── flash_x2.safetensors
└── RelayVSR_X4/
    ├── wan.safetensors
    └── flash.safetensors

Each directory is discovered automatically and its scale is detected from the safetensors metadata. The relayvsr key is registered with ComfyUI, so you can also point other disks at it through extra_model_paths.yaml. After adding or moving folders, restart ComfyUI (or use the loader's custom path field).

Nodes

RelayVSR Model Loader — category model/loaders

| Input | Description | | --- | --- | | bundle | Directory under models/relayvsr containing a matched Wan/Flash pair | | dtype | bf16 (default) or fp16 for the floating-point weights | | offload | Move Wan blocks to the GPU one at a time. Enabled by default; disable it on 32 GB+ cards for faster reference passes | | path | Advanced: absolute path to a weight directory, overrides bundle |

Output: RELAYVSR_MODEL (connect to the upscale node).

RelayVSR Upscale — category image/upscaling

| Input | Default | Description | | --- | --- | --- | | model | — | Output of the loader | | images | — | IMAGE batch; each frame is processed as part of one sequence | | sr_ratio | 2 | Output scale. 2× weights require 2; 4× weights accept 2 or 4 | | chunk_len | 16 | Frames per streaming window. Larger chunks are faster but use more VRAM | | seed | 0 | Only affects the initial noise of the reference frames |

Output: IMAGE at height * sr_ratio × width * sr_ratio. Progress and cancellation are supported.

INT8 conv-rotation weights

convert_to_int8_convrot.py rewrites the Wan backbone into ComfyUI's native INT8 tensor-wise format with conv rotation. Only the attention/FFN linear layers inside blocks.* are quantized; convolutions, norms, embeddings, biases and proj_out stay in full precision. The Flash decoder is copied unchanged. The loader detects <module>.comfy_quant automatically and routes F.linear through ComfyUI's INT8 kernel.

Run it with ComfyUI's Python, since the script imports torch and comfy_kitchen (both come with ComfyUI). From the portable root (ComfyUI_windows_portable) on Windows:

python_embeded\python.exe ComfyUI\custom_nodes\Comfyui-Relay-VSR\convert_to_int8_convrot.py ComfyUI\models\relayvsr\RelayVSR_X2

For a manual install, use the Python from your ComfyUI virtual environment instead. The script requires CUDA (--device cuda is the default).

The output goes to RelayVSR_X2_int8_convrot/ by default and is selected as a separate bundle:

ComfyUI/models/relayvsr/
├── RelayVSR_X2_int8_convrot/
│   ├── wan_x2_int8_convrot.safetensors
│   └── flash_x2.safetensors

INT8 bundles support the same block offload as BF16 bundles. The converter currently targets the 2× filenames (wan_x2.safetensors / flash_x2.safetensors).

Notes

• Wan in BF16 is about 9.4 GB of weights; INT8 conv-rotation roughly halves that. The Flash decoder is about 120 MB.
• A 768×1344 clip with 16-frame windows needs about 18–19 GB in BF16. On 24 GB cards, leave offload enabled: the Wan KV cache alone is about 12 GB at 1080p, and holding the Wan weights in VRAM as well will run out of memory.
• There is no prompt input: RelayVSR uses a fixed prompt embedding.
• The model is 1-step distilled (sigma 1 → 0, no CFG), so seed has very limited influence.

License and attribution

This project is licensed under the Apache License 2.0.

It contains code ported from:

• RelayVSR — https://github.com/kopperx/RelayVSR (Apache-2.0)
• LightX2V — https://github.com/ModelTC/LightX2V (Apache-2.0)

Modifications: removed the LightX2V framework/registry dependencies, rewrote attention, KV caching and streaming-window orchestration for ComfyUI, and added ComfyUI nodes, INT8 loading and model-directory handling.

RelayVSR, LightX2V and their logos belong to their respective owners; this project is not endorsed by or affiliated with them. Model weights are not distributed here and may be governed by separate terms on Hugging Face / GitHub.