// the find
kijai/ComfyUI-SUPIR
SUPIR upscaling wrapper for ComfyUI
A ComfyUI custom node wrapping the SUPIR image upscaler, which runs SDXL as an img2img pipeline with a specialized ControlNet for restoration. It exposes the research model's sampler controls in ComfyUI's node graph and supports fp8, tiled VAE, and LoRA loading. The README now says it's superseded by native ComfyUI core support — this wrapper is in maintenance-only mode.
The decomposed multi-node design is better than the original single-node approach — separating the first stage, sampling, and VAE decoding lets you slot in other preprocessing without fighting the wrapper. fp8 support for the UNet means you can run 512→2048 upscaling on 10GB VRAM, which is a real constraint for most hobbyists. Tiled VAE is built in and works around memory spikes at high resolutions. The author also ships pruned safetensors models instead of requiring the original multi-file checkpoint mess.
The project is explicitly abandoned — the README's first heading is 'FINAL update', and SUPIR is now in ComfyUI core, so anyone starting fresh has no reason to use this wrapper. Hardware requirements are punishing: 10GB VRAM for a 512→1024 upscale, potentially 32GB+ RAM, which rules out anything other than a high-end workstation. Non-commercial-only license buried at the bottom of the README is a landmine for anyone who might try to use this in a production pipeline. Video upscaling processes frames one by one with no batching or temporal consistency, making it useless for anything longer than a few seconds.