Freedom251 › Image Upscaling Software
SUPIR
by SUPIR
What SUPIR is
SUPIR is a free, open-source image restoration project founded in 2024. It develops a high-fidelity general restoration model built on a large diffusion generative prior for enhancing compressed, noisy, or blurry photos. Text prompts and configurable inference settings support image processing and upscaling, with examples covering landscape, face, animal, gaming, cinematic, and vintage images. The repository provides SUPIR-v0Q with default training settings and SUPIR-v0F trained with light degradation settings. Command-line runs accept input and output folders and let users choose an upsampling ratio, model, random seed, minimum output size, sampling steps, prompts, and color-fixing choices. Local setup uses Python 3.8, Conda, pip, and the requirements file, plus SDXL and LLaVA model dependencies. A linked SupPixel AI service offers online processing and upscaling. The project documents a lower-VRAM mode using 12 GB for diffusion and 16 GB for LLaVA, with slower performance. Software use, reproduction, and distribution are strictly non-commercial unless written permission is obtained.
Who it suits
SUPIR suits developers and image users who can run a local, open-source restoration workflow or use the linked online app. It is relevant for restoring and enlarging varied photo types with prompt and inference controls.
What it does well
- Open-source and free to use
- Supports batch processing and custom scale
- Offers two documented model variants
- Provides prompt and sampling controls
- Links to an online processing tool
What to know first
- Commercial use requires prior written permission
- Local setup needs several dependent models
- Lower-VRAM mode runs more slowly
- Requires substantial documented memory allocations
Verdict
SUPIR provides configurable restoration and upscaling for users comfortable with model-based workflows. Its non-commercial restriction, dependencies, and hardware requirements should be checked before deployment.
Facts from SUPIR's own pages
- Batch processing
- Yes · github.com, 22 Sept 2026
- Commercial use
- The repository declares SUPIR software available strictly for non-commercial purposes unless prior written permission is obtained from Dr. Jinjin Gu. · github.com, 12 Oct 2026
- Commercial-use restriction
- The repository declares SUPIR software use, reproduction, and distribution strictly non-commercial unless prior written permission is obtained from Dr. Jinjin Gu. · github.com, 12 Oct 2026
- Compute requirements
- The repository notes a lower-VRAM configuration that uses 12G for diffusion and 16G for LLaVA, with slower performance. · github.com, 12 Oct 2026
- Dependent models
- The documented dependencies include SDXL CLIP encoders, SDXL base, LLaVA CLIP, and LLaVA v1.5 13B. · github.com, 12 Oct 2026
- Founded
- 2024 · github.com, 28 Sept 2026
- Hardware requirements
- The repository describes a lower-VRAM demo mode as using 12 GB for diffusion and 16 GB for LLaVA, with slower processing. · github.com, 12 Oct 2026
- Image controls
- Inference options include an upsampling ratio, model selection, random seed, minimum output size, sampling steps, prompts, and color-fixing choices. · github.com, 12 Oct 2026
- Image enhancement
- The project presents SUPIR for image processing and upscaling, including compressed, noisy, or blurry images. · xpixel.group, 12 Oct 2026
- Installation
- The repository instructions create a Python 3.8 Conda environment and install dependencies from requirements.txt. · github.com, 12 Oct 2026
- Local installation
- The repository instructions use Python 3.8, Conda, pip, and the project's requirements file to install SUPIR. · github.com, 12 Oct 2026
- Maximum scale
- custom · github.com, 22 Sept 2026
- Model
- SUPIR is described as a high-fidelity general image restoration model based on a large-scale diffusion generative prior. · supir.xpixel.group, 12 Oct 2026
- Model choices
- The repository provides SUPIR-v0Q with default training settings and SUPIR-v0F trained with light degradation settings. · github.com, 12 Oct 2026
- Model dependencies
- The installation instructions list SDXL CLIP encoders, an SDXL base checkpoint, and LLaVA CLIP and v1.5 13B among dependent models. · github.com, 12 Oct 2026
- Model variants
- The repository provides SUPIR-v0Q with default training settings and SUPIR-v0F trained with light degradation settings. · github.com, 12 Oct 2026
- Online app
- The repository links to an online app and describes SupPixel AI as an easy-to-use image processing and upscaling tool built on SUPIR. · github.com, 12 Oct 2026
- Online demo
- The repository links to SupPixel AI, an online image processing and upscaling tool built on SUPIR. · github.com, 12 Oct 2026
- Purpose
- SUPIR aims to develop practical algorithms for photo-realistic image restoration in the wild. · github.com, 12 Oct 2026
- Quality and fidelity
- The repository says its empirical hyperparameters affect image quality and appearance, and describes a tradeoff between visual quality and fidelity. · github.com, 12 Oct 2026
- Restoration approach
- The project describes SUPIR as a high-fidelity general image restoration model based on a large-scale diffusion generative prior. · supir.xpixel.group, 12 Oct 2026
- Run options
- The repository supports command-line inference with input and output folders, an upsampling ratio, model selection, and configurable sampling and prompt parameters. · github.com, 12 Oct 2026
- Security and compliance
- The project pages opened do not state security certifications or compliance claims. · github.com, 12 Oct 2026
- Support
- The repository directs questions to [email protected] or [email protected]. · github.com, 12 Oct 2026
Where SUPIR stands on Freedom251
- Image Upscaling Software#84 of 99