Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

Carol-lyh/Jeveration

Machine Learning

Run a frozen vision-language model that generates full images in one forward pass across text-to-image, image-to-image and dense prediction tasks.

119
9
Python
Apache-2.0
4 days old

What it does

Jeveration is an inference codebase for a frozen vision-language model that outputs a complete 512 × 512 image as one joint decision in a single forward pass. The authors report 78 ms per text-to-image request on one NVIDIA H200. A single checkpoint covers text-to-image, image-to-image (structure-to-image, colorization, style transfer) and dense prediction (segmentation, depth, surface normals). The project is aimed at researchers and developers who want to reproduce or build on the paper Jeveration: System One Models Now Generate; a project page, weights and an online demo are linked from the README.

Setup

You need Python 3.12 and a CUDA GPU. Install the Python dependencies and download the 6.5 GB weights into a local folder:

pip install -r requirements.txt
hf download linlinlinlinsdada/Jeveration --local-dir weights

The backbone (Qwen3.5-4B), the TA-TiTok tokenizer and CLIP's text tower are fetched from the Hugging Face Hub on first use.

Quick start

From the repository folder, run text-to-image with the downloaded checkpoints:

python generate.py --weights weights/jeveration.pt --decoder weights/decoder512.pt \
    --prompt "A stone cottage with a thatched roof and a flower garden in the countryside." --seed 0 --out samples/

Image-to-image uses --task to select a mode. For example, hed or depth takes an edge or depth map plus a --caption describing the wanted image; watercolor, anime and ghibli take a photo:

python image_to_image.py --weights weights/jeveration.pt --source-decoder weights/source_decoder512.pt \
    --task watercolor --image examples/photo_snowcabin.jpg --out watercolor.png

Dense prediction returns segmentation, depth and surface normals; it was trained on NYUv2 indoor scenes:

python dense_predict.py --weights weights/jeveration.pt --image examples/room_1.png --out room_dense/

A Python pipeline is also available for running all tasks from one object (README usage section), and python app.py ... starts a Gradio web demo with examples for the three tasks. Every script lists its options with --help. A fixed seed gives the same image on the same hardware and software.

Constraints

  • Requires a CUDA GPU and Python 3.12; runtime downloads include the Qwen3.5-4B backbone, TA-TiTok tokenizer and CLIP text tower, so internet access is needed on first run.
  • The dense prediction model was trained on NYUv2 indoor scenes, so accuracy on other domains is not established by the materials.
  • Image-to-image modes have specific input expectations: hed/depth need an edge or depth map (or a photo with --from-photo), colorize treats input as grey, and anime/ghibli/watercolor take a photo.
  • Code and Jeveration weights are Apache-2.0; third-party code and models downloaded at runtime keep their own licenses (THIRD_PARTY_NOTICES.md).

HotRepo's take

Jeveration appeared on the new-project list for Python on 2026-11-22, gaining 3 stars in the observed one-hour window (101 to 104 stars) with 7 forks. That is a short observation window, so the record supports only that the project drew early attention right after release, not a trend judgment. The value here is the single-checkpoint, single-pass design for generation and dense prediction; whether it reproduces the reported 78 ms latency depends on the H200 hardware used in the paper, which the repository's own materials do not establish for other GPUs.

Written by AI from the project's public information and HotRepo's records, updated . The project's own docs are the source of truth.

You might also like

  1. OpenMOSS/MOSS-ImageAn open 5B diffusion transformer for text-to-image that keeps text legible and lets you control element placement.Python1,752
  2. yyfz/Pi4Reconstructs 4D scene geometry and camera motion from images or video in a single forward pass, for graphics and vision researchers.Python118