🌏 中文版
Model Info
| Field | Value |
|---|---|
| Model ID | briaai/Fibo-2.0-dev (plus a 4-step distilled variant, briaai/Fibo-2.0-edit-turbo) |
| Vendor | Bria AI |
| Parameters | 8B (flow-matching diffusion transformer) |
| Context Window | N/A (image model; trained natively at roughly 1 megapixel, recommended sizes are multiples of 16 such as 1024×1024, 1424×800, 1248×832) |
| Price per image (fal.ai) | $0.04 |
| Price per image (Bria API, per Artificial Analysis testing) | $0.032 ($32 / 1,000 images) |
| Open source | Yes (CC BY-NC 4.0, non-commercial only; commercial use requires a separate agreement with Bria) |
| Release date | 2026-10-20 |
| Official announcement | Hugging Face Model Card |
| HuggingFace | briaai/Fibo-2.0-dev |
| Family | FIBO (Bria's structured-prompt image model family; preceded by FIBO 1.0 in 2025 and Fibo-Edit-1.5 in August 2026) |
Capability Highlights
- The first open-weights model to put image generation and instruction-based editing in a single checkpoint: one model covers generation, masked editing, multi-reference editing with up to four images, and an "Inspire" mode that borrows a reference's style without copying it
- Ranks #1 on Artificial Analysis's open-weights image-editing leaderboard (1087 Elo, from 4,819 human blind-preference votes) and #2 on text-to-image — statistically tied with the leader, Qwen-Image-2.1 (around 1083±13 Elo)
- Costs $0.032–0.04 per image (Bria API / fal.ai), undercutting Grok Imagine Image Quality ($0.05/image), Qwen Image 2.0 Pro ($0.075/image), and Luma UNI 1 Max ($0.10/image)
- Carries forward FIBO's VGL (Visual GenAI Language) design: structured JSON prompts replace natural-language descriptions, so a single attribute — lighting, camera angle, an outfit — can be changed without disturbing the rest of the scene
Benchmark Results
| Metric | Fibo 2.0 Dev | Predecessor (FIBO 1.0 / Fibo-Edit-1.5) | Best competitor |
|---|---|---|---|
| AA-Image-Editing v2.1 (open-weights rank) | #1 (1087 Elo) | No precise Elo published; Bria says "every editing action improved" | Qwen-Image-2.1, ~1083±13 Elo (tied within margin of error) |
| AA-Image-T2I v2.1 (open-weights rank) | #2 | Same as above | Qwen-Image-2.1 (open-weights #1) |
| Enhancement/restoration editing (mixed with paid models) | #6 | - | - |
| Text-to-image — human anatomy accuracy | #18 | - | - |
| Text-to-image — text rendering / layout | #19–21 range | Bria calls this one of the biggest gains this generation | Leading models (e.g., the Gemini 2.5 Flash Image line) remain clearly ahead |
⚠️ These Elo scores and rankings are cited from Bria's own model card, which in turn cites Artificial Analysis (an independent third-party evaluator) — they are not Bria's self-reported benchmarks. But third-party outlet ArtRealmAI checked the public leaderboard around launch and found Fibo 2.0 hadn't been listed yet; Qwen-Image-2.1 still showed 1083±13 Elo at the top. Bria's claimed lead is only about 4 Elo points, within the confidence interval — meaning the "#1/#2" claim is a self-reported figure made ahead of the leaderboard's formal update, pending independent reproduction.
Comparison with Predecessors/Competitors
Compared with FIBO 1.0 and the August 2026 Fibo-Edit-1.5, the biggest architectural shift is merging two previously separate model lines — generation and editing — into a single 8B checkpoint. Teams used to deploy a generation model and an editing model separately; now one deployment covers both, and Bria reports gains on every editing action and text-to-image capability over the prior generation.
Against the current open-weights text-to-image leader, Qwen-Image-2.1, the two are essentially tied on raw Elo (the gap falls within the margin of error). Fibo 2.0's real differentiator isn't raw image quality — it's the controllability and reproducibility that come from structured JSON prompts: an "edit" becomes a diff against a JSON field rather than a fresh natural-language prompt and a roll of the dice. The weak spots are just as clear: text rendering and complex layout still sit in the #19–21 range, well behind the top closed models.
On pricing, Fibo 2.0's $0.032–0.04 per image undercuts the field — cheaper than Grok Imagine, Qwen Image 2.0 Pro, and Luma UNI 1 Max. But the license takes a different path than most rivals' "open weights means you can ship it" stance: these weights are non-commercial only, and commercial use needs a separate agreement with Bria, so enterprises need to check the license scope before deploying.
What This Means for Agent Development
The structured JSON (VGL) prompt format is the most meaningful piece here for agent architecture: an "edit" is defined as a localized change to a readable, diffable, version-controllable JSON blueprint, rather than handing a black-box model a new paragraph of natural language and hoping. That means visual generation work can be pulled into the same diff/review/reproduce workflow as code.
- If you're building an agent for e-commerce assets or brand visual consistency: you can have the agent manipulate JSON fields (lighting, camera angle, outfit color) instead of rewriting prompts, which produces more predictable output that's easier to A/B test
- If you need one pipeline that covers both generation and editing — virtual try-on, product compositing, multi-reference composition — Fibo 2.0 folds both needs into a single deployment, cutting the cost of maintaining two separate models
- Not a fit: scenarios that need commercial rights without the budget for a Bria agreement (the open weights are non-commercial only), or scenarios that need precise text rendering or complex layout — an area where it still clearly trails top models
Today's Takeaway
I used to assume "image generation" and "image editing" were necessarily two different models with two separate inference pipelines, each optimized on its own. Fibo 2.0 folding both into one checkpoint shows that boundary isn't a technical necessity — it's a product-segmentation leftover. That's worth keeping in mind when designing multimodal agent tool interfaces: tools don't have to be split along "generate vs. edit." They can instead be split along something closer to the actual data shape — whether the input includes an image.