Bria AI drops Fibo 1.5 with JSON-native prompts and 4-step generation

Fibo 1.5 is an 8B-parameter open-source text-to-image and editing model from Bria AI. Instead of relying on prompt poetry, it is natively built to follow structured JSON inputs.

What it can do

While you can still feed it everyday text prompts, Fibo 1.5 truly shines when paired with structured JSON schemas spanning over 1,000 words.

You can break an entire image down into distinct JSON keys: description, location, relative size, shape, color, texture, camera lens, and lighting. The model reads the schema and renders the exact scene you specified.

It runs in three primary modes. In Generate mode, an integrated vision-language model takes your brief thought, expands it into a clean JSON schema, and outputs both the image and the generated schema.

In Refine mode, it provides true disentangled editing. If you tell it to add a blanket draped over the armchair or switch to an 85 mm lens, it only changes that specific attribute while preserving the rest of the composition.

In Inspire mode, you drop in an existing image, let the model extract the structured breakdown, and remix it with your creative spin.

Reviewers testing the checkpoint noted that while raw aesthetic richness might not beat heavyweight closed models, its prompt adherence and ability to iteratively stack edits—like adding books on the floor, then a cat on the chair—is remarkably rock-solid.

Under the hood and how to run it

Under the hood, Fibo 1.5 is an 8B DiT model using flow matching. It pairs a SmolLM3-3B text encoder with a Wan 2.2 VAE, driven by an upstream VLM like a tuned Qwen-2.5 or Gemini 2.5 Flash.

The real upgrade in version 1.5 is a two-stage distillation process: DMD collapses sampling steps, and DMD-R polishes realism.

The distilled model needs only 4 to 6 inference steps without classifier-free guidance, meaning you can skip negative prompts entirely and enjoy massive latency cuts.

Bria trained the entire system exclusively on fully licensed data, ensuring complete compliance with GDPR and the EU AI Act.

The model weights and inference code are open for non-commercial use on Hugging Face, with commercial licenses available separately. You can run it via Diffusers, ComfyUI, Draw Things, and DiffusionBee, or hook into cloud APIs like Fal.ai and Replicate.

Keep in mind that this specific distilled checkpoint is locked in for fast 4-to-6-step inference and is not meant to be fine-tuned; base model weights are required if you plan to train custom checkpoints.

Why it matters

If you just want a generic wallpaper, standard consumer generators will do. But if you build production pipelines or product graphics, Fibo 1.5 offers the exact predictability traditional models lack.

Replacing fuzzy natural language with deterministic JSON turns image generation into code you can version-control and edit systematically. Combined with clean legal provenance and sub-second generation speeds, we think it sets a sensible blueprint for enterprise-grade generative design.