BlurLabsAI / blurlabs

# Blur Labs Open Imaging Models
> Foundation models for computational imaging from Blur Labs: one network that reconstructs images from the measurements of any
> imaging system, with no retraining per task.
## Reconstruct Anything Models (RAM)
RAM solves general imaging problems of the form `y = N(Ax)`, where `A` models the deterministic acquisition model, 
and `N` some noise degradation. To solve this problem, RAM takes as 
inputs i) the measurements `y`, ii) the acquisition physics `A` and iii) the noise level(s), and returns
the reconstructed image. 
RAM covers any 2D imaging problem, from photography, microscopy to medical imaging (MRI, CT, ...), and any other problem
the forward operator of which can be written with DeepInverse, that provides
a large collection of imaging operators: denoising, deblurring,
demosaicing, super-resolution, PET, ultrasound, inpainting, (multi-coil) MRI, CT, and many more.
This repository holds all open-source versions of the model. **RAM 2** is the latest version: since the first Reconstruct Anything Model
(ICLR 2026), we have improved the architecture and training significantly.
RAM 2 outperforms RAM on all imaging tasks, and per task, performs better than or close to state-of-the-art specialized models _trained specifically and only on one task_. See the blog post for a summary of results.
### Versions
| Version | Class | Parameters | Paper / results | Weights | License |
|-----------|------------------------|---|------------------------------------------------------------------------------------------------------------|-----------------------------------------------------|---|
| **RAM 2** | `blurlabs.models.RAM2` | 359M | blog post | Hugging Face | GPL-3.0-only. For alternative commercial license, contact Blur Labs |
| RAM 1 | `deepinv.models.RAM` | 36M | ICLR'26 paper, project page | Hugging Face | BSD-3-Clause |
## Installation
Install the latest release directly from GitHub:
 [code] 
For an editable source installation with the example dependencies:
 [code] 
The RAM 2 weights are gated. First request access on HuggingFace, then run
`hf auth login` or set the `HF_TOKEN` environment variable. Weights are downloaded on first use.
## Quick start
 [code] 
Notes:
- 2D images are represented as tensors of shape `B,C,H,W` , with `C=1` (grayscale), `C=2` (real and imaginary parts, e.g. MRI) or `C=3` (RGB) channels, of any size.
- Noise levels are read from `physics.noise_model`, or given with the keywords `sigma` (Gaussian), `gain` (Poisson)
 and `l` (Gamma). `sigma` given alone is used as the exact noise standard deviation, scalar or per pixel.
- With `physics=None`, RAM 2 is a denoiser: `model(y, sigma=0.1)`.
### Examples
Install the example dependencies with `pip install -e ".[examples]"`. Model weights and example data are
downloaded on first use.
Below, we provide a few examples of usage of RAM 2 on various image reconstruction problems.
> **Tip**:
> 
> RAM 2 can be used as a drop-in replacement for any existin