Fanghua-Yu/SUPIR
Source: https://github.com/fanghua-yu/supir
SeedVR/SeedVR2 (https://github.com/ByteDance-Seed/SeedVR, Apache 2.0 per README, one 80GB H100 for roughly 100 frames at 720x1280) was fetched as a possible alternative but not recorded in this lane.
Structured fields
{
"archived": false,
"checked_at": "2026-10-17T00:00:00.000Z",
"usefulness": "CVPR 2024 diffusion-based photo-realistic restoration/upscaling built on SDXL with LLaVA captioning for guidance. High-fidelity but heavy; its custom 'SUPIR Software License Agreement' with a non-commercial declaration is the clearest example in this lane of a code release that cannot be assumed usable in commercial pipelines.",
"source_type": "repository",
"availability": "available",
"dependencies": [
"SDXL base 1.0",
"SDXL CLIP encoders",
"LLaVA v1.5 13B + LLaVA CLIP",
"Python 3.8 conda env"
],
"source_owner": "Fanghua Yu / SupPixel Pty Ltd",
"next_check_at": "2026-10-31T00:00:00.000Z",
"hardware_notes": "Sourced: low-VRAM option about 12G (diffusion) + 16G (LLaVA); default uses two GPUs.",
"known_limitations": [
"Non-commercial restriction.",
"Multiple large downloads, legacy Python 3.8 env, heavy VRAM.",
"Diffusion upscaling can alter faces and fine details; the quality/fidelity trade-off is documented by the authors."
],
"upstream_updated_at": "2025-05-12T09:49:11.000Z"
}Evidence claims
- [sourced_fact] README: restoration/upscaling code for a CVPR 2024 paper; low-VRAM mode (half params, tiled VAE, 8-bit LLaVA) described as needing about 12G for the diffusion model and 16G for LLaVA; default examples use two GPUs. source
- [sourced_fact] A 'Non-Commercial Use Only Declaration' limits use, reproduction and distribution to non-commercial purposes, with commercial use needing written permission; no separate licence stated for checkpoints SUPIR-v0Q/v0F or dependencies (SDXL, LLaVA v1.5 13B, optional Juggernaut-XL). source
- [sourced_fact] LICENSE file is titled 'SUPIR Software License Agreement', copyright 2024 SupPixel Pty Ltd, work jointly owned by SupPixel Pty Ltd. source
- [sourced_fact] README notes a quality-vs-fidelity trade-off: settings that raise visual quality can lower fidelity to the input. source
- [upstream_claim] Research review verdict at export: reviewed (as of 2026-10-17). Recorded by the research pipeline, not a hub review outcome; evidence.review stays unreviewed until an independent reviewer records one.
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| code | SUPIR-Software-License-Agreement | restricted | restricted | restricted | unknown | Custom agreement plus Non-Commercial Use Only Declaration; commercial use needs permission from the named author. Only the first lines of the LICENSE were read. |
| model_weights | unknown | unknown | unknown | unknown | unknown | Checkpoint licence not stated on README; dependent SDXL (OpenRAIL++-M), LLaVA and Juggernaut models carry separate terms not evaluated here. |