SeFi-Image¶

Development line: project:sefi-image · thread sefi-image-development
Last event: 2026-06-26 · 1 dated since 2026-06-26 · Researched: 2026-09-05 · confidence: high

What it is¶

SeFi-Image separates semantic and texture latent streams so structure denoises ahead of detail. It offers 1B, 2B, and 5B Base models, a 5B RL model, and 1B/2B/5B Turbo models. Base and RL use 50 steps at guidance 4.0; Turbo defaults to four steps at guidance 1.0. We run it for controlled local research, not commercial deployment or hosted inference.

Development line¶

• 2026-06-26 — SeFi-Image official project resources were linked. On 2026-06-26, a development-line message linked the project website, source repository, and Hugging Face page. The evidence confirms these links, but shows no specific release, model, or technical milestone.

What changed¶

2026-06-26 — arXiv v3 documented SeFi-Image as a public semantic-first-diffusion T2I family with 1B, 2B, and 5B scales and DMD2-distilled Turbo variants.

How to use this¶

As of 2026-06-26, we use the linked project website, GitHub repository, and Hugging Face page to evaluate SeFi-Image.

1. Accept the checkpoint access conditions, then choose a checkpoint: Base for analysis or fine-tuning, RL for alignment-oriented generation, or Turbo for fast generation. — https://huggingface.co/SeFi-Image/SeFi-Image-5B-Base
2. Install the repository runtime in Python 3.11 with PyTorch matching local CUDA, then install Diffusers, Transformers, Accelerate, Safetensors, Hugging Face Hub, OmegaConf, and Pillow. — https://github.com/jmliu206/SeFi-Image
3. Run inference.py with a checkpoint ID, prompt, output directory, and seed; use a Base checkpoint such as SeFi-Image/SeFi-Image-5B-Base for the default 50-step path. — https://github.com/jmliu206/SeFi-Image
4. For low-latency generation, select a Turbo checkpoint and set four steps with guidance scale 1.0. — https://github.com/jmliu206/SeFi-Image
5. For a standard Diffusers workflow, load a -diffusers checkpoint with SeFiPipeline or DiffusionPipeline in BF16 on CUDA. — https://huggingface.co/SeFi-Image/SeFi-Image-5B-turbo-diffusers

Best practices¶

• Keep Turbo at 4, 8, or 10 denoising steps and guidance scale 1.0; the project documents four steps as its default. — https://github.com/jmliu206/SeFi-Image
• Use Base checkpoints as the starting point for fine-tuning and analysis; reserve Turbo for fast generation and RL for stronger alignment-oriented output. — https://huggingface.co/SeFi-Image/SeFi-Image-5B-Base
• Treat the released weights as CC BY-NC 4.0, gated, research-use material; apply human oversight, moderation, validation, and compliance checks before any broader use. — https://github.com/jmliu206/SeFi-Image

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The sources do not provide a dated changelog describing the substantive differences between arXiv v3, v4, and v5.
• No first-party source establishes a production-serving option; reviewed Hugging Face Diffusers cards state that no Inference Provider deploys the models.

Sources¶

Agent brief¶

• Subject: project:sefi-image, thread sefi-image-development, 1 dated events 2026-06-26 → 2026-06-26.
• Practical note: As of 2026-06-26, use the linked project website, GitHub repository, and Hugging Face page as starting points to evaluate SeFi-Image.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.