InvSR
Arbitrary-steps image super-resolution via diffusion inversion for high-res outputs.
Enhance real-world images and AIGC outputs using large pre-trained diffusion models with 1 to 5 steps.
InvSR is a Python-based image super-resolution tool featured at CVPR 2025 that uses a deep noise predictor and partial noise prediction strategy. It is designed for graphic designers and creative professionals who need to upscale low-quality images and enhance AIGC content quickly.
What it is
It leverages pre-trained diffusion models like SD-Turbo to perform real-world image super-resolution and AIGC image enhancement. The framework supports flexible sampling between one and five steps, allowing you to scale up images from 1k to 4k using specific chopping options.
Who it's for
It suits graphic designers and digital artists with technical skills who want to harness diffusion priors for upscale tasks locally or via provided demos. It does not suit non-technical users who require a standard point-and-click desktop application without handling code, Python environments, or GPUs.
Key features
Partial Noise Prediction
Constructs an intermediate state of the diffusion model as the starting sampling point using a deep noise predictor.
Flexible Sampling Steps
Supports an arbitrary number of sampling steps ranging from one to five to balance speed and quality.
Large Image Chopping
Includes chopping options like chopping size and batch size to handle large 1k to 4k image upscaling on limited GPU memory.
Color Fixing Options
Provides wavelet color fixing to reproduce consistent quantitative results on test datasets.
Up and running in 3 steps
Our honest verdict
What's great
- Achieves superior super-resolution performance even with a single sampling step
- Provides an online Replicate and Hugging Face demo for quick testing without local setup
- Includes Docker compose support for containerized deployment
- Offers training scripts to train or fine-tune your own noise predictors with custom datasets
Where it falls short
- Requires significant technical setup including Python 3.10, PyTorch 2.4.0, xformers, and a suitable CUDA-capable GPU
- Running large models locally demands substantial GPU memory and hardware resources
- Training or fine-tuning requires preparing specific LPIPS models and multi-GPU configurations
Pricing & license
The software is free under its repository licence, and the only financial cost is running it yourself on your own hardware or using third-party hosted demo links provided in the project.
Open remote AI-training roles
767 open remote AI-training roles, updated November 14, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can I run InvSR with limited GPU memory?
Yes, you can add the chopping size and chopping batch size options to manage memory constraints when processing large images.
How many steps does the sampling process require?
The method supports an arbitrary number of sampling steps ranging from one to five, where even a single step yields strong results.
Are there online demos available?
Yes, the README links to both a Replicate demo and a Hugging Face space for immediate testing.
