Bria Fibo 2.0 in ComfyUI: Generate and Edit Images With the 8B Open Model

Bria released Fibo 2.0 this week, and it is one of the few open weights models that does both image generation and image editing with a single set of weights. It is an 8B parameter diffusion transformer that reads structured JSON prompts, edits photos from plain instructions or masks, takes up to four reference images, and ships with a 4-step Edit Turbo checkpoint for near-instant edits. ComfyUI merged native support on October 21, 2026. This guide explains what Fibo 2.0 is, which checkpoint to download, the files and folders for ComfyUI, the settings that matter, how the JSON “VGL” prompts work, and the license catch you need to know before you use it for client work.

What Is Bria Fibo 2.0?

Bria is an Israeli AI company that built its business on image models trained only on licensed data, which is why its models are popular with brands and e-commerce teams. Its first FIBO model was known for extremely long, structured captions that described lighting, camera, composition and color as separate fields. Fibo 2.0 keeps that structured approach and adds editing to the same model.

The weights were uploaded to Hugging Face on October 18 and 19, 2026. Here is what is inside:

The design is clever. Instead of feeding thousands of raw text tokens into the image model, Fibo 2.0 runs your prompt through Qwen3-VL and then squeezes the result into a few hundred summary tokens with a small Perceiver. Five of the DiT blocks also look directly at internal layers of the vision language model. For edits, your input images go through both the VLM, so the model understands what is in them, and the VAE, so it can keep the pixels you did not ask to change.

If you already run FLUX 3 Image Klein or MOSS-Image, you already have the FLUX.2 VAE it needs.

Dev vs Edit Turbo: Which Checkpoint?

Edit Turbo is distilled from Dev. Four steps without CFG means roughly 25 times fewer model evaluations than Dev’s 50 steps with CFG, so edits that take a minute with Dev take seconds with Turbo. My advice: install both. Use Edit Turbo for everyday edits like background swaps, color changes and product placement, and use Dev when you want to generate from scratch or when Turbo’s edit is not quite right.

Fibo 2.0: model evaluations per imageDev runs 50 steps with CFG (2 passes per step). Edit Turbo runs 4 steps, no CFG.Dev (generate + edit)100Edit Turbo (edit only)4Fewer evaluations = faster. Source: Bria model cards via ComfyUI Wiki, Oct 20 2026.

Hardware: VRAM, RAM and Disk

Bria does not publish VRAM requirements. What we know is the size of the files: the bf16 transformer is 16.25 GB, the Qwen3-VL-4B encoder is 8.88 GB in bf16 and the VAE is 0.34 GB, about 25.5 GB per checkpoint. The Comfy-Org repack adds INT8 ConvRot builds of both transformers, which roughly halves the transformer size.

Those are my planning estimates based on file sizes and how ComfyUI offloads the text encoder after encoding, not measured numbers. Multi-reference edits use more memory because each reference image adds tokens.

Plan for 32 GB of system RAM and about 30 GB of free disk space if you want both INT8 checkpoints plus the encoder and VAE.

Files to Download and Where They Go

The easiest source is the Comfy-Org/Fibo-2.0 repack, which already uses ComfyUI’s folder layout. The original Bria repos are gated: log in to Hugging Face and accept Bria’s terms on the model page before you can download either checkpoint.

Check the repo’s file list for exact file names, since repacks are sometimes renamed in the first week.

Install Fibo 2.0 in ComfyUI

1. Update ComfyUI. Native support arrived in PR #17240 on October 21, 2026. Update to the latest development build: run update/update_comfyui.bat on the portable build, or git pull on a git install. Desktop users may need the nightly channel until the next release.
2. Accept Bria’s terms on briaai/Fibo-2.0-dev and briaai/Fibo-2.0-edit-turbo if you download from Bria directly.
3. Download the transformer(s), the Qwen3-VL-4B encoder and the FLUX.2 VAE into the folders above.
4. Restart ComfyUI and build the graph: Load Diffusion Model for the transformer, CLIPLoader with type fibo2, the TextEncodeFibo2 node for your prompt and any reference images, Load VAE, a sampler and VAE Decode.
5. Test text-to-image with Dev at 1024 x 1024, then try an edit with Turbo.

Two alternatives exist. The community node pack akhaliq/Fibo-2.0-ComfyUI loads the same files from models/diffusion_models. And ComfyUI’s built-in Bria nodes and the ComfyUI-BRIA-API pack call Bria’s hosted API instead, which costs money but needs no GPU.

Recommended Settings

Good starting sizes are 1024 x 1024, 1152 x 896, 896 x 1152, 1344 x 768 and 768 x 1344. For edits, resize your source photo to around 1 megapixel first, keeping both sides divisible by 16, so the model does not have to guess how to reframe it.

Structured JSON Prompts (VGL) Explained

Fibo 2.0 was trained on VGL, short for Visual GenAI Language. It is a JSON-style description of an image with separate fields for things like the subject, lighting, composition, color palette, camera and depth of field. The original FIBO captions ran past 1,000 words.

You do not have to write VGL. Plain prompts work, and the TextEncodeFibo2 node handles both. But structured prompts are where Fibo shines, because you can change one field (say, the lighting) and leave everything else identical. The pipeline also returns the JSON alongside the image, so you can save it, replay it and compare two versions line by line, which is great for product photography where consistency matters.

Bria’s bria-vgl Python package has two helpers: create() turns a short prompt into a full VGL document, and modify() turns an edit instruction into VGL changes. Note that these helpers call Google’s Gemini and need a GEMINI_API_KEY, so they are not fully local. If privacy matters, write your VGL by hand or use a local LLM to expand your prompts.

A simplified example of the idea (field names are illustrative, not the full schema):

{
  "subject": "a ceramic coffee mug with a matte sage-green glaze",
  "setting": "wooden kitchen counter, morning",
  "lighting": "soft window light from the left, gentle shadows",
  "camera": "50mm, eye level, shallow depth of field",
  "color_palette": "warm neutrals with sage green accents"
}

Editing: Instructions, Masks and Multi-Reference

This is the main reason to try Fibo 2.0 over a generation-only model.

Instruction edits

Give it a photo and a sentence like “change the jacket to red leather” or “make it sunset”. Bria says the model is trained to keep identity, layout, lighting and fine detail outside the target region. Edit Turbo is built for exactly this.

Masked edits

Paint a mask over the area you want changed and only that region is regenerated. Use this when an instruction edit touches more than you wanted, for example when changing a product label without the model “improving” the background.

Multi-reference edits

You can pass up to four images and refer to them in the prompt as <image_1>, <image_2> and so on. Bria lists virtual try-on, product placement and appearance transfer as target uses. For example: “Place the sneaker from <image_2> on the person in <image_1>, matching the lighting.”

Bria has also announced, but not yet shipped, a multi-turn editing LoRA, native background removal, a product shot LoRA, a “boring reality” style LoRA, mask-based object erasing and image expansion. Expect more editing workflows over the next few weeks.

How Good Is It? Benchmarks

Bria’s model card cites an Artificial Analysis snapshot from October 20, 2026: first among open weights models on the image editing arena and second on the text-to-image arena. An independent write-up from ExplainX pointed out that on editing, Fibo 2.0’s Elo (1038) and Qwen-Image 2.1’s (1041) are within each other’s error bars, so “tied for the lead” is the fairer reading. Either way, it is in the top group of open editing models right now.

Bria also claims better text rendering, textures and anatomy than the first FIBO, and the best prompt adherence among open models. Those are vendor claims, so test your own use case.

The License: CC BY-NC 4.0

This is the part people skip and regret. Fibo 2.0’s weights are released under Creative Commons Attribution-NonCommercial 4.0. You can use them for personal projects, research, learning and experiments. You cannot use them for commercial work, which includes client jobs, product photos for your store or a paid app, without a separate agreement with Bria.

If you need commercial use, Bria sells API access (reported at about $32 per 1,000 images, though I could not confirm that against Bria’s own pricing page) and commercial licenses. Bria’s selling point is that its training data is licensed, which matters to brands worried about copyright. That claim has not been independently audited, but it is the reason many companies pick Bria.

For commercial local use, an Apache 2.0 model like MOSS-Image is the safer choice, though it cannot edit images.

Fibo 2.0 vs FLUX 3 Klein vs MOSS-Image

All three share the FLUX.2 VAE, so trying them side by side costs only the transformer and encoder downloads. Pick Fibo 2.0 for the best open editing quality on non-commercial projects, Klein for fast editing you can use commercially under its revenue threshold, and MOSS-Image for an Apache-licensed generator.

Troubleshooting

No “fibo2” option in CLIPLoader

Your ComfyUI predates PR #17240. Update to the latest development build and restart.

Download fails with 401 or 403

The Bria repos are gated. Log in, accept the terms on each model page, and use a Hugging Face token with read access.

Edit Turbo ignores my negative prompt

That is expected. Turbo runs at guidance 1.0 and does not use negatives. Rephrase the positive instruction instead.

Edit Turbo produces junk from a text prompt alone

Turbo is editing only. Use Dev for text-to-image.

Out of memory with multiple references

Use the INT8 transformer, reduce each reference to about 1 megapixel, and use fewer references. Start ComfyUI with --lowvram on 12 GB cards.

Diffusers says BriaFibo2Pipeline does not exist

The Diffusers pipeline is still in a pull request (#14991). Install the branch build named on the model card, or use ComfyUI.

FAQ

Can I use Fibo 2.0 commercially?

Not with the open weights. They are CC BY-NC 4.0. Commercial use needs a Bria license or the paid API.

Is Fibo 2.0 open source?

It is open weights: you can download and run it. The non-commercial license means it is not open source in the usual sense.

How much VRAM does Fibo 2.0 need?

Bria has not said. With the INT8 transformer, a 16 GB card is a comfortable target, and 12 GB should work with offloading.

Does it need an internet connection?

No. ComfyUI runs it fully offline. Only Bria’s optional bria-vgl prompt helpers call Gemini online.

Can I train LoRAs on Fibo 2.0?

Yes, Bria says Dev is the base meant for LoRAs and full fine-tunes. Training tools are still catching up in the first week.

Can I try it without installing anything?

Yes. Bria runs free Hugging Face Spaces: briaai/Fibo-2.0-T2I for generation and briaai/Fibo-2.0-editing for edits with Turbo.

Final Thoughts

Fibo 2.0 is the strongest open weights image editor you can run locally this week, and the 4-step Edit Turbo makes it practical on a 16 GB card. Update ComfyUI past PR #17240, grab the INT8 Edit Turbo and Dev transformers from the Comfy-Org repack, reuse your FLUX.2 VAE, and start with simple instruction edits before moving to masks and multi-reference. Just keep the non-commercial license in mind before you put it to paid work.

More local image AI guides:

• MOSS-Image in ComfyUI, an Apache 2.0 generator with JSON layout control.
• FLUX 3 Image Klein in ComfyUI, the other new open editor.
• Local Background Removal with U2NETP on CPU, until Fibo’s background removal ships.
• Best Local AI Video Models in 2026, to animate your edited images.

Sources

• ComfyUI Wiki: Fibo 2.0, Bria’s 8B open weights model
• ExplainX: Fibo 2.0 claims vs verified
• Hugging Face: briaai/FIBO (license and commercial access)