Fibo-2.0-dev Image Generation Model: 32GB+ VRAM, File List

At a Glance
| Item | Value |
|---|---|
| Repository | briaai/Fibo-2.0-dev |
| Family guide | Fibo-2.0-dev guide (3 articles) |
| Published | 2026-10-18 |
| License | bria-fibo |
| Formats | safetensors |
| Access | Gated on Hugging Face (license agreement required) |
| Paper | arXiv:2511.06876 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Bria AI has released “briaai/Fibo-2.0-dev", an image generation model that supports long, structured captions.
The model supports a variety of visual tasks, including text-to-image, image-to-image, image editing, inpainting, and multi-reference.
It is part of the FIBO series of models based on the related research paper “Generating an Image From 1,000 Words: Enhancing Text-to-Image With Structured Captions" (arXiv:2511.06876), and is developed with the aim of eliminating ambiguities in generation that rely on traditional short prompts and achieving high controllability.
Specifications
The specifications confirmed in the documentation and related papers are as follows:
- Number of Parameters: 8.1B (approx. 8.1 billion parameters).
- Input/Output & Functions: Supports text-to-image, image-to-image, image editing, inpainting, and multi-reference.
- Prompt Processing Method: Takes long structured captions annotated with the same set of detailed attributes as input.
- Integration Mechanism (Architectural Features): Employs “DimFusion", a fusion mechanism that integrates intermediate tokens from a lightweight LLM without increasing the overall token length.
- Implementation & Pipeline: Compatible with Hugging Face’s Diffusers library, using
BriaFibo2Pipelineto execute inference. - Supported Languages: Supports English (en).
- Distribution Format: Provided in safetensors format.
- License Terms: The license category is “bria-fibo" (license:other in the metadata), and repository access is set to a gated state requiring an application for agreement on Hugging Face.
Performance and Quality
According to the research paper upon which this model is based (arXiv:2511.06876), FIBO is reported to have achieved state-of-the-art prompt alignment among open-source text-to-image models.
Many conventional text-to-image models are trained to generate detailed images from short prompts, which poses a challenge where the model arbitrarily fills in missing detailed information, making precise control intended by the user difficult. The underlying research of Fibo-2.0 uses long structured captions with a unified, fine-grained attribute set applied to all training data, maximizing expressive coverage while enabling disentangled control over individual visual elements.
Additionally, a new evaluation protocol called “Text-as-a-Bottleneck Reconstruction (TaBR)" has been introduced for quality assessment. This is a method to verify how accurately real images can be reconstructed through a loop of caption generation and image generation. It is stated to directly measure the controllability and expressiveness of generative models even for extremely long prompts that were difficult to measure with existing evaluation metrics.
Note that since individual benchmark score tables are not listed in the model card of the public repository, specific numerical comparisons await future official announcements and community verification.
Strengths and Use Cases
briaai/Fibo-2.0-dev excels at controlling various visual elements within an image with high precision using detailed, long structured prompts.
With conventional image generation models, when short prompts are entered, the model tends to automatically complement missing details, making it difficult to specify fine-grained compositions and attributes as desired. In contrast, this model enables disentangled control to manipulate each element independently by processing structured captions annotated with the same fine-grained attribute set. Therefore, it is suited for professional image production and design applications where you want to faithfully reflect intended layouts and attributes without relying on accidental depictions.
Furthermore, this model is tagged for a wide range of editing and reference tasks in addition to simple text-to-image:
- Image-to-image: Suitable for using an existing image as input and changing the style or content while maintaining the structure.
- Image editing and inpainting: Supports tasks to modify and complement specified parts of already generated or existing images.
- Multi-reference: Utilizes multiple reference images as input for generation that maintains consistency of characters or objects, and advanced workflows combining multiple visual elements.
- Fine-tuning: Expected to be used as a base model for additional training tailored to specific domains or custom attribute systems.
The supported language is English (en), and it demonstrates its strengths in scenarios where long attribute instructions written in English are accurately interpreted to construct images.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 8.1B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 32GB (RTX 5090, etc.) | BF16 | 23.7GB | 28.5GB |
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
The publisher distributes this model as safetensors.
License — bria-fibo: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.
Access: You must accept the license on Hugging Face before downloading.
Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Our Own Measurements
Measurements We Did Not Take
We have not confirmed that this site meets the commercial-use terms of this model’s license (bria-fibo). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Recent Models in the Same Size Class
Models with 4–15B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| Qwen/Qwen-Image-2.1-Turbo | 7.1B | 48GB | qwen-research | Qwen-Image-2.1-Turbo 8-Step Fast Image Generation Model: 48GB+ VRAM (2026-11-17) |
| lodestones/Kromoss | 5.1B | 32GB | apache-2.0 | Kromoss 1024px Image Generation Model: 32GB+ VRAM, File List (2026-11-17) |
| unsloth/MOSS-Image-SFT-GGUF | 5.1B | 32GB | apache-2.0 | MOSS-Image-SFT-GGUF 5.1B Text-to-Image Generation Model: 32GB+ VRAM (2026-11-17) |
| Viggle/Qwen-Image-2.1-viggle-turbo | 7.1B | 48GB | qwen-research | Qwen-Image-2.1-viggle-turbo Image Generation Model: 48GB+ VRAM (2026-09-23) |
Distributed Files
Weight files published in briaai/Fibo-2.0-dev, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
text_encoder/model.safetensors |
8.88GB |
transformer/diffusion_pytorch_model-*-of-00002.safetensors |
16.25GB (2 split files) |
vae/diffusion_pytorch_model.safetensors |
336MB |
How to Get It
briaai/Fibo-2.0-dev is available on the Hugging Face model repository.
Weight files are provided in safetensors format, and it supports Hugging Face’s Diffusers as a library. The pipeline is loaded using the dedicated BriaFibo2Pipeline.
When obtaining this model, attention should be paid to the following points:
- Necessity of License Agreement (Gated Model): This model is set to a gated state. Therefore, before downloading and using the files, you must complete the agreement procedure for the license terms on the Hugging Face model page.
- License System: The license is set to “bria-fibo".
The repository and related publication pages are as follows:
- Model Repository: briaai/Fibo-2.0-dev
- Research Paper: arXiv:2511.06876
Related Articles
- Fibo-2.0-edit-turbo Image Editing Model: 32GB+ VRAM, File List
- Fibo-2.0 Image Generation and Editing Diffusion Model: 32GB+ VRAM
- Black Forest Labs Releases FLUX.3 Image Klein
- Qwen-Image-2.1-Turbo 8-Step Fast Image Generation Model: 48GB+ VRAM
What to Read Next
- Find models by VRAM (This model needs at least 32GB) → VRAM quick reference
- Explore the same model family → Fibo-2.0-dev family overview (3 articles)
- Formats this model is available in → Safetensors format guide and models
- Other models for the same task → Other image generation models

