文件清单 线路:镜像(hf-mirror.com) · 切换
下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。
说明
Fibo 2.0
作者:CivitaiOfficial
Fibo 2.0 is an 8B-parameter image generation and editing model from Bria AI. It is trained on structured captions of up to 1,000+ words, so a short prompt or a plain-language edit instruction is expanded into a detailed structured description before the image is drawn. That gives it fine control over composition, lighting, color and text, and lets an edit change one thing while keeping the rest of the image intact.
Originally released by Bria AI on Hugging Face. All credit for the model goes to the Bria AI team. Civitai hosts it so creators can run it on-site - head to the original repos for the full model cards, updates, and Bria's own tools.
Built by
- Bria AI
- Paper authors: Eyal Gutflaish, Eliran Kachlon, Hezi Zisman, Tal Hacham, Nimrod Sarid, Alexander Visheratin, Saar Huberman, Gal Davidi, Guy Bukchin, Kfir Goldberg and others - Generating an Image From 1,000 Words: Enhancing Text-To-Image With Structured Captions
Versions on Civitai
- Dev - the full, non-distilled base model. Civitai uses it for text-to-image. Bria recommends 50 steps at guidance 5.0, at about one megapixel (1024x1024, 1424x800 or 1248x832). Fine-tunes and LoRAs are trained against Dev.
- Edit Turbo - distilled from Dev with the same architecture, for editing only. Bria recommends 4 steps at guidance 1.0; a negative prompt has no effect. It is not intended for fine-tuning.
Both are the Comfy-Org repackaging of Bria's weights, in bf16 and an int8 build whose outputs differ slightly from the original.
Architecture
A flow-matching model in the latent space of the FLUX.2 VAE. Prompts are encoded by Qwen3-VL-4B-Instruct, and a small Perceiver resampler compresses its output into a few hundred tokens for a 44-block single-stream DiT, with some of the VLM's internal layers feeding the DiT directly through cross-attention.
Editing
Give Edit Turbo one or more images and an instruction such as "make it night time" or "replace the mug with a glass". The instruction is turned into a structured edit, so the change stays local and the rest of the image is preserved.
Content and licence
Generation with Fibo 2.0 on Civitai is limited to PG-13 content. Bria releases the weights under CC BY-NC 4.0 for non-commercial use; see the licence linked on this page for how it applies here.
Links
- Fibo 2.0 Dev: huggingface.co/briaai/Fibo-2.0-dev
- Fibo 2.0 Edit Turbo: huggingface.co/briaai/Fibo-2.0-edit-turbo
- Paper: arXiv 2511.06876
- Bria platform: platform.bria.ai
- Bria Discord: discord.com/invite/Nxe9YW9zHS
