Generating an image from 1,000 words. Very excited to release Fibo , the first ever open-source model trained exclusively on long, structured captions. Fibo sets a new standard for controllability and disentanglement in image generation [1/6]

Short textual descriptions don’t capture the rich details of high-quality images, so models trained on short captions are hard to control. Instead, we train on long and structured captions, ensuring each one captures all necessary information in a clear, organized way. This results in a highly expressive model with incredible prompt-following. [2/6]

But it gets even better, although the model wasn’t trained for editing (no image pairs during training), native disentanglement emerges from our structured representation, enabling iterative refinement. [3/6]

Of course, writing 1,000 words prompt is hard. And so, we use a VLM to complete additional details. This empower our model with additional world knowledge and reasoning. We use Gemini in default, but we also release a fine-tuned version of Qwen-VLM for the open source and research communities. [4/6]

Fibo is open-source for non-commercial use. Try it now: Model Weights: https://huggingface.co/briaai/FIBO Playground: https://platform.bria.ai/labs/fibo A full technical paper with all training details is coming very soon - stay tuned [5/6]

On a personal note, I find it incredibly powerful to simply view the structured prompt JSON - it makes these chaotic models more explainable and approachable. This model is definitely underexplored. It behaves differently, and there’s so much to learn. I can’t wait to see what the community achieves with the model and I’d love your feedback [6/6]