The Pulse
Bria Releases Fibo 2.0, an 8B Model for Image Editing and Generation
Bria has released Fibo 2.0, an 8-billion-parameter model that generates and edits images using one set of weights. Its weights are available for non-commercial use under CC BY-NC 4.0; commercial use requires a separate agreement with Bria.
AI.info Team · Written with AI · fact-checked ·
One set of weights for two jobs
Bria released Fibo 2.0 on October 20, an 8-billion-parameter model designed to generate images and edit existing ones with the same base checkpoint. The company’s Hugging Face model card presents the model as the next generation of its FIBO family, with support for text-to-image generation, instruction-based edits, masks and multiple reference images.
That combination targets a practical split in image workflows: one model produces a new image from a prompt, while another changes a supplied image without disturbing everything around the edit. Fibo 2.0 is built to handle both, including local edits confined by a mask and compositions that draw on as many as four reference images. Bria says it was trained exclusively on licensed data.
The release comes in two versions. Fibo 2.0 Dev is the full, non-distilled base for generation, editing and fine-tuning; Edit Turbo is a distilled variant aimed at faster editing, with Bria specifying four inference steps and no classifier-free guidance. The distinction matters for users choosing between a model intended for experimentation and adaptation and one tuned for shorter editing runs.
Structured prompts shape the workflow
Fibo 2.0 relies on Bria’s Visual GenAI Language, or VGL: a structured format that represents details such as lighting, composition, camera settings and objects in JSON. The model card says its training used long, structured captions, with individual examples reaching more than 1,000 words. Bria’s premise is that spelling out visual attributes makes it easier to change one part of an image without prompting unrelated details to drift.
The card describes an 8B flow-matching model operating in the latent space of the FLUX.2 VAE. It names Qwen3-VL-4B-Instruct as the prompt encoder and describes a single-stream transformer with 44 blocks. For users who do not want to write structured prompts themselves, Bria’s bria-vgl package turns short requests into VGL; the model card says that helper runs on Gemini and requires a Gemini API key.
Developers can use the Dev checkpoint through Diffusers, while Bria also offers generation and editing through its own API and lists ComfyUI support. Bria’s release documentation says existing API integrations can keep the same code while the image-generation and editing endpoints use Fibo 2. The model card’s local-inference examples, by contrast, require the relevant weights and software dependencies.
“Open weights” comes with a commercial limit
Bria makes the weights available under Creative Commons Attribution–NonCommercial 4.0, or CC BY-NC 4.0. The Hugging Face page gates access behind a form and account sign-in, and specifies that commercial use requires a separate agreement with Bria. The release therefore offers access to model weights, but not an unrestricted commercial license.
That qualification will shape whether companies can use the model for in-house prototyping, client work or products for sale. Bria says its training data is licensed, but that claim is separate from the license governing use of the released weights. Organizations considering deployment still need to assess both the model terms and the hardware and software needed to run it.
Leaderboard ranks are a snapshot, not a product guarantee
Bria’s model card reports that, in its October 20 snapshot of Artificial Analysis, Fibo 2.0 ranked first among open-weight models for image editing and second for text-to-image generation. Artificial Analysis says its image boards rank models by Elo scores derived from pairwise human-preference votes. Those standings provide a comparative signal from that evaluation, not a guarantee of results on every prompt, editing task or production workflow.
Bria also lists follow-on work as forthcoming rather than available at launch: multi-turn local editing, background removal, a product-shot LoRA, erasing by mask and image expansion. The released Dev weights provide the base model now; those additional editing tools remain on Bria’s roadmap.