VSRDecode - ComfyUI Built-in Node Documentation - ComfyUI

VSR Decode turns the keyframe latents from RelayVSR Conditioning back into a full video with the RelayVSR Flash decoder. The decoder adds its detail on top of a bicubic upscale of the low-quality frames you pass in, at the decoder’s x2 or x4 factor. Frame count stays the same.

Inputs

Parameter	Description	Data Type	Required	Range
`vae`	The RelayVSR Flash decoder, loaded with Load VAE: `relayvsr_flash_vae_x2_fp32.safetensors` or `relayvsr_flash_vae_x4_fp32.safetensors`. The file’s metadata sets the upscale factor, so it has to match the model.	VAE	Yes	N/A
`samples`	The latent returned by the sampler, carrying one frame per keyframe.	LATENT	Yes	N/A
`images`	The same low-quality video given to RelayVSR Conditioning.	IMAGE	Yes	N/A
`detail`	How much of the decoder’s detail is kept on top of the bicubic upscale: 1 is the full model output, 0 is the plain upscale. (default: 1.0)	FLOAT	Yes	0.0 - 1.0 (step 0.01)

Outputs

Output Name	Description	Data Type
`images`	The upscaled video with every frame restored.	IMAGE

Notes

Pass the same low-quality video to both nodes. The decoder reads it frame by frame, so a mismatch in length or resolution gives a wrong or broken result.
Lowering detail softens the restored detail and keeps the output closer to a plain bicubic upscale, which can help on very noisy sources.

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): 946ac04bc9f205fd6d1088554caaa16238086e570f212e747b166e3fb0d9b3c5