Fibo Edit

This node runs Bria's FIBO-Edit, an 8-billion-parameter instruction-editing model, directly in ComfyUI. FIBO belongs to the same 2025–26 generation as Qwen-Image-Edit, Flux Kontext, and Flux 2 Klein - the edit models that quietly demoted mask-inpainting to a fallback for most jobs. Where FIBO breaks rank is its input format: it's trained on long, structured JSON captions (think 1,000-word annotated essays, not a sentence), and that's simultaneously its superpower and its biggest usability sin. The community liked the output well enough - "least slopped new image model of 2025" is about the highest praise a 2025 model could get - but nobody wants to hand-write that JSON.

That's why this pack ships two nodes. The FiboEdit node you're looking at is the drawing half: it takes an image, an optional mask, and a structured JSON prompt, then runs the diffusion transformer. Its sibling, FiboEdit_VLM, writes that JSON from plain English. They're designed to be chained - the example workflow is literally VLM → FiboEdit → SaveImage - and you can't get far without both.

How it works

The pack vendors Bria's own BriaFiboEditPipeline rather than calling any API. Under the hood it's the now-familiar modern stack: the 8B transformer (MMDiT-style), a Wan VAE, a SmolLM3-based text encoder that digests the structured captions, and a flow-matching scheduler - the same design family as Flux. On first run the node downloads whatever model you name into ComfyUI/models/fibo_edit/, not your checkpoints folder, so ComfyUI Manager won't track it and the first run is a long, silent download. The pipeline runs with CPU offload enabled, so the whole thing doesn't need to sit in VRAM at once.

One rule is enforced: json_prompt must be valid JSON containing an edit_instruction key, or the node raises. That's the trap. You can't type "make the armor red" and get away with it - the model genuinely wants the structured caption.

The inputs that actually matter

• json_prompt - the only one that'll bite you. Valid JSON with an edit_instruction. Write it by hand or, far saner, feed it from FiboEdit_VLM.
• image - the image you're editing. mask (optional) - restrict the edit to a region; a fully-zero mask is silently ignored, so if nothing seems to change, check your mask.
• model - defaults to boredcoder/Fibo-Edit-8bit, an 8-bit community quant (~16 GB download). The official briaai/Fibo-Edit is ~40 GB and gated on Hugging Face (more below).
• steps (default 50), cfg (default 5), seed (-1 = random).
• low_vram - here's the thing: it doesn't offload anything, it just forces cfg to 1. The author's own description says so. It'll cut VRAM a bit and change the look; don't treat it as a magic toggle.

Output is an IMAGE - wire it to a Preview or Save node like anything else.

Installing it

Via ComfyUI Manager, search for the pack (its title is "Fibo Edit Node for ComfyUI"). Or, manually:

cd ComfyUI/custom_nodes
git clone https://github.com/boredcoderyt/ComfyUI-Fibo-Edit
cd ComfyUI-Fibo-Edit
pip install -r requirements.txt

The requirements include transformers>=4.57.1, diffusers>=0.36.0, bitsandbytes, accelerate, and qwen-vl-utils. That diffusers floor is genuinely new - if your ComfyUI install is old, expect an upgrade before this runs.

Where people get burned

• The official model is gated. briaai/Fibo-Edit requires a Hugging Face login plus accepting the license. If auto-download fails, run huggingface-cli login. The default 8-bit quant repo isn't gated, which is exactly why it's the default.
• The example workflow's model string is stale. It references boredcoder/fibo_edit_4bit, which doesn't resolve for anonymous users. Use the README's boredcoder/Fibo-Edit-4bit instead (4-bit, ~12 GB).
• That 40 GB full-precision repo includes a leftover transformer_old folder you don't need. The quant repos skip it.
• FIBO-Edit is CC-BY-NC-4.0 - fine for personal play, not for commercial work.

The honest bottom line: if you want a sentence-driven edit model, Qwen-Image-Edit or Klein get you there with less ceremony. Reach for FIBO when you want the structured-caption control it was actually built for - and use the VLM node to write the JSON.