PRX

PRX is Photoroom's open-source text-to-image model, a 1.3B transformer released as nine checkpoints at 256, 512 and 1024 px. PRX Pixel is a 7B sibling that denoises pixels directly, with no autoencoder. LibreYOLO runs all ten through the LibreGen factory on the diffusers pipelines.

Tasks
image generation
Sizes
prxpixel: 7b at 1024 px
Install
pip install "libreyolo[gen]"
Support tier
Sibling tier, since v1.8.0. A separate product surface with its own factory and contract.
Upstream
PRX and PRX Pixel by Photoroom, Apache-2.0. Paper, source
Licenses
Code MIT, weights Apache-2.0 (PRX's T5Gemma text encoder: Gemma Terms of Use). Commercial use

Install

bash
pip install "libreyolo[gen]"

PRX Pixel needs diffusers>=0.39 and says so when it loads; the extra's floor is 0.38 so it can sit next to the Marigold extra.

Generate

Weights download from LibreYOLO's byte-identical mirrors of Photoroom's repositories on first use.

Python
from libreyolo import LibreGen model = LibreGen("prx-512-sft")result = model("a red fox sleeping in fresh snow", seed=0)result.save("fox.png")             # the PNG also stores prompt, seed and params
Settings
from libreyolo import LibreGen model = LibreGen("prx-512-sft-distilled")          # 8-step distillationresult = model("a red fox", steps=8, guidance=1.0) model = LibreGen("prx")                            # prx-1024-betaresult = model("a red fox", negative="blurry", imgsz=1024, seed=0) model = LibreGen("prxpixel", offload=True)         # prxpixel-7b, needs diffusers>=0.39results = model(["a red fox", "a blue jay"], n=2, save=True)
CLI
libreyolo generate prompt="a red fox" model=prx-512-sft negative="blurry"libreyolo generate prompt="a red fox" model=prxpixel-7b seed=0 n=4

A string is always a prompt. One prompt with n=1 returns a Results; a list of prompts or n > 1 returns a flat list, and stream=True a generator. result.generated holds the image (array, image), the prompt, the seed that reproduces it and the params in effect. Every family knob is optional; one the family does not have raises and lists the supported ones. See the image generation task page for prompts, seeds, devices and offload.

Variants

ModelParametersKnobs (defaults)GPU memory (library figure)
prx-1024-beta (alias prx)1.3Bsteps (28), guidance (5.0), negative ("")about 21 GB
prx-512-sft-distilled1.3Bsteps (8), guidance (1.0), negative ("")about 16 GB
prx-512-sft, prx-5121.3Bsteps (28), guidance (5.0), negative ("")about 16 GB
prx-512-dc-ae-sft-distilled1.3Bsteps (8), guidance (1.0), negative ("")about 16 GB
prx-512-dc-ae-sft, prx-512-dc-ae1.3Bsteps (28), guidance (5.0), negative ("")about 16 GB
prx-256-sft, prx-2561.3Bsteps (28), guidance (5.0), negative ("")about 16 GB
prxpixel-7b (alias prxpixel)7Bsteps (28), guidance (5.0), negative ("")about 18 GB

The number is the native resolution. -dc-ae uses the 32x DC-AE autoencoder instead of the FLUX.1 VAE, so its sides round up to a multiple of 32 instead of 16. -sft is the supervised fine-tune, ready to use, and -sft-distilled the 8-step distillation of it. Checkpoints without -sft are base models meant for fine-tuning. prx-1024-beta is Photoroom's 1024 preview. PRX Pixel bundles a Qwen3-VL text encoder and has no autoencoder.

Export

TaskONNXTorchScriptExecuTorchTensorRTOpenVINOPaddleMNNRKNNAxelerancnnTFLiteCoreMLCore AI
Image generationImage generation to ONNX: not supportedImage generation to TorchScript: not supportedImage generation to ExecuTorch: not supportedImage generation to TensorRT: not supportedImage generation to OpenVINO: not supportedImage generation to Paddle: not supportedImage generation to MNN: not supportedImage generation to RKNN: not supportedImage generation to Axelera: not supportedImage generation to ncnn: not supportedImage generation to TFLite: not supportedImage generation to CoreML: not supportedImage generation to Core AI: not supported

export(), train() and val() raise, as for every LibreGen family.

Licensing

Check the license on the Hugging Face repository of the specific weights you download. Every checkpoint in the LibreYOLO org carries one, and they are not always the same across a family. That repository is the authoritative source; the summary below describes what applied when this page was last verified.

This is a description of the licenses involved, not legal advice. If the answer matters commercially, read the licenses yourself and take your own counsel.

Original work
PRX and PRX Pixel, Photoroom
Upstream license
Apache-2.0
LibreYOLO code
MIT
Weights
Apache-2.0 (PRX's T5Gemma text encoder: Gemma Terms of Use), republished at huggingface.co/LibreYOLO
Interpretation
Photoroom releases the PRX and PRX Pixel weights under Apache-2.0, a permissive license, so they can be used in commercial and closed-source products; redistributing them means keeping the license text and notices, and the license grants a patent license. The exception is PRX's bundled T5Gemma 2B text encoder, which stays under Google's Gemma Terms of Use and the Gemma Prohibited Use Policy: commercial use is allowed, but the prohibited uses pass on to anyone you share it with, and PRX's model cards forbid those uses for the whole model. PRX's autoencoder is the FLUX.1 [schnell] VAE (Apache-2.0) or DC-AE (MIT); PRX Pixel has no autoencoder and bundles the Qwen3-VL-Embedding-2B text encoder (Apache-2.0). LibreYOLO's adapter is its own MIT code that calls the Apache-2.0 diffusers PRXPipeline and PRXPixelPipeline; no upstream source is ported. Weights download from LibreYOLO's byte-identical mirrors of Photoroom's nine PRX repositories and prxpixel-t2i (LibreYOLO/LibrePRX<size> and LibreYOLO/LibrePRXPixel7b), each pinned to a commit and not gated.

PRX and PRX Pixel are by Photoroom. The weights are Apache-2.0, so commercial use is fine. The exception is PRX's bundled T5Gemma 2B text encoder, which stays under Google's Gemma Terms of Use and Prohibited Use Policy: commercial use is allowed, but the prohibited uses pass on to anyone you share it with, and PRX's model cards forbid them for the whole model. The autoencoder is the FLUX.1 [schnell] VAE (Apache-2.0) or DC-AE (MIT); PRX Pixel's Qwen3-VL-Embedding-2B text encoder is Apache-2.0.

LibreYOLO's adapter is its own MIT code on the Apache-2.0 diffusers PRXPipeline and PRXPixelPipeline. Weights download from LibreYOLO/LibrePRX<size> and LibreYOLO/LibrePRXPixel7b, byte-identical mirrors pinned to a commit, not gated.

Citation

@misc{photoroom2026prx,
  title  = {PRX: an open-source text-to-image model},
  author = {Bertoin, David and Frigg, Roman and Almazán, Jon and Andres, Eliot},
  year   = {2026},
  url    = {https://huggingface.co/blog/Photoroom/prx-open-source-t2i-model},
}

Copied from the authors' citation block at github.com/Photoroom/PRX#citation.

Verified against LibreYOLO v1.8.0. Support tables, checkpoints and benchmark numbers on this page are generated from the released library and the published weights, not written by hand.