Generating a new image from a prompt is one task, while editing an existing image without disturbing its identity, composition, lighting, or fine details is another.Even capable open image models can struggle with precise prompt adherence, readable text, realistic textures, and correct human anatomy.
Another challenge is controllability, users need to describe exactly what they want while preserving everything they did not ask to change.
FIBO 2.0 model (Dev and Edit Turbo) released by BriaAI focus on addressing these issues by combining image generation and editing in one unified model while retaining the structured prompting approach that made the original FIBO family controllable.
 |
| Fibo 2.0 Model architecture (ref-official research paper) |
It is the next generation of Bria's FIBO model family. It brings text-to-image generation, instruction-based image editing, and iterative refinement into a unified system. The model builds on Visual GenAI Language (VGL), which uses long, structured captions to help connect detailed instructions with visual results.
FIBO 2.0 adds improvements in general-purpose editing, in-image text rendering, material textures, and human anatomy. One of its key distinctions is its training data. Like other FIBO releases, FIBO 2.0 is trained exclusively on licensed data, making data provenance an important part of its approach to open image generation. You can get more detailed insights by accessing their research paper. The models is registered under non commercial international license, means you may use it for non commercial use only. You have to grant agreement for using it commercially.
Basically two models variants released-
(a) FIBO 2.0 Dev- It is the full, non-distilled base model. It uses classifier-free guidance to provide greater control over generation quality and is intended for text-to-image inference, professional fine-tuning, and research. Choose Dev when you want to generate images from prompts or customize the model for a specific workflow.
(b) FIBO 2.0 Edit Turbo- It is the distilled, production-ready editing variant. It performs image edits in just four steps without classifier-free guidance, reducing inference latency and cost. Its purpose is focused: fast image editing rather than text-to-image generation. It is the practical option when you want to apply instructions to existing images quickly.
Installation
1. Make sure you have ComfyUI already installed. If already have, just update to its latest version from the Manager.
2. Download Fibo2 models for ComfyUI compatability. choose any of them as per your requirements and resources. for low steps with faster inference and editing use the edit turbo model and use dev (non-distilled base) variant for fine-tuning.
-fibo2_dev_bf16.safetensors
-fibo2_dev_int8_convrot.safetensors
-fibo2_edit_turbo_bf16.safetensors
-fibo2_edit_turbo_int8_convrot.safetensors
Save any of them into ComfyUI/models/diffusion_models folder.
3. Download text encoders models.
-qwen3vl_4b_bf16.safetensors
-qwen3vl_4b_int8_convrot.safetensors
Save any of them into ComfyUI/models/text_encoders folder.
4. Download fibo turbo lora (fibo_2.0_turbo_lora_avg_rank_270_bf16.safetensors). Its for faster inference rank 270 lora model. Save it into ComfyUI/models/loras folder.
5. Download vae (flux2-vae.safetensors) model. Save it inside ComfyUI/models/vae folder. Its the same vae that you already have so downloading again isn't required.
6. Restart and refresh ComfyUI.
Workflow
1. Download Fibo 2.0 Dev and edit turbo workflows from hugging face repository.
2. Drag and drop into ComfyUI.
3. Load the models (diffusion models, vae and text encoders) into its relevant nodes.
4. Set Ksampler settings.
Fibo 2 Dev -
Steps-50
CFG-5.0
Resolution supported- 1024×1024 (the default 1:1), 1424×800 (16:9) or 1248×832 (3:2)
Fibo 2 Edit turbo -
Steps-4
CFG-1.0
Maximum 4 images supported to do editing in the format <image1>, <image2> so on.
5. Put Prompts into Prompt box.
6. Hit Run to start genration.
