The eight-billion-parameter FIBO family is trained on licensed data and detailed structured captions. Version 1.5 applies DMD followed by DMD-R refinement to generate in four to six steps without classifier-free guidance. A vision-language model can expand short instructions or image references into the structured prompt.
Fibo 1.5 supports generation, targeted refinement, and image-inspired creation through local inference, ComfyUI, and hosted integrations. Weights are gated for noncommercial access, while commercial use requires separate licensing. The distilled checkpoint is not intended for fine-tuning; users needing adaptation should start from the base model.
