README.md

---
title: SeFi-Image
sdk: gradio
sdk_version: 6.19.0
app_file: app.py
python_version: 3.12
license: cc-by-nc-4.0
short_description: SeFi-Image Base, Turbo, and 5B RL ZeroGPU demo.
models:
  - SeFi-Image/SeFi-Image-1B-Base
  - SeFi-Image/SeFi-Image-2B-Base
  - SeFi-Image/SeFi-Image-5B-Base
  - SeFi-Image/SeFi-Image-5B-RL
  - SeFi-Image/SeFi-Image-1B-turbo
  - SeFi-Image/SeFi-Image-2B-turbo
  - SeFi-Image/SeFi-Image-5B-turbo
---

# SeFi-Image

Public Gradio ZeroGPU Space for SeFi-Image 1B, 2B, and 5B Base/RL/Turbo checkpoints.

- Models on Hugging Face: https://huggingface.co/SeFi-Image
- Project page: https://jmliu206.github.io/sefi-web/

Base checkpoints use the 50-step guidance-4.0 setting. Turbo checkpoints use
the 4-step guidance-1.0 setting. The 5B RL checkpoint is the reinforced-learning
variant of the 5B model family.

The model defaults match the SeFi model cards:

| Family | Models | Steps | Guidance |
| --- | --- | ---: | ---: |
| Base | 1B, 2B, 5B | 50 | 4.0 |
| RL | 5B | 50 | 4.0 |
| Turbo | 1B, 2B, 5B | 4 | 1.0 |

The SeFi checkpoints are gated under CC BY-NC 4.0. The Space uses an `HF_TOKEN`
secret to download model files, so the Space owner account must first be approved
for each selected SeFi model repository.

The Space generates fixed 1024x1024 images. The included examples were generated
from this Space with `SeFi-Image-5B-RL`. Selecting an example fills the prompt,
parameters, seed, and output image while leaving the randomized seed option
enabled for the next generation.

The examples cover anime-realistic portraiture, realistic editorial portraiture,
cinematic fantasy landscapes, and complex poster or packaging prompts with
Cyrillic, Korean, Chinese, and mixed-script typography.

The vendored `sefi/` inference package comes from
`jmliu206/SeFi-Image` at commit `2f02744a187639ee41a296f8177cbbe7e5f333f5`
and is covered by `LICENSE-SEFI-INFERENCE`.