SeFi‑Image: Semantic‑First Diffusion Model Family and ComfyUI Support Question
Prepare custom inference scripts for SeFi‑Image or await ComfyUI integration, noting its dual VAE and CC BY‑NC 4.0 license.
Develop custom inference scripts for SeFi‑Image or monitor for ComfyUI integration.
Summary
SeFi‑Image, a Semantic‑First Diffusion model family, introduces a novel architecture that combines a semantic latent with a texture latent and employs a dual VAE, including a baked‑in VAE. The family includes 1 b, 2 b, 5 b, and 5 b RL variants, all using the flux 2 VAE and Qwen3 VL 2b/4b for text encoding. The model is released under a strict CC BY‑NC 4.0 license, limiting commercial use. The arXiv paper demonstrates impressive generation quality, but the community questions whether native ComfyUI support will be added or if it will remain locked behind experimental Python scripts. The author notes that the model’s architecture is unconventional, with a dual VAE and semantic‑first diffusion. The paper also highlights the use of flux 2 VAE across all sizes, which may simplify integration. The lack of native ComfyUI support could hinder adoption among users who rely on that interface. The discussion invites developers to explore custom inference scripts or await potential future integration.
Key changes
- Dual VAE architecture with baked‑in VAE
- Uses flux 2 VAE across all sizes
- Qwen3 VL 2b/4b for text encoding
- Semantic‑first diffusion (semantic + texture latent)
- CC BY‑NC 4.0 license (non‑commercial)
- No native ComfyUI support currently
- Architecture unconventional for community
- Requires custom inference scripts or future integration