Fibo-2.0-ComfyUI

Run FIBO 2.0 locally in ComfyUI — no Bria API token, no per-call cost. The official ComfyUI-BRIA-API nodes call Bria's hosted API; these nodes run the open weights on your own GPU.

Now with native nodes (fibo2_native/): the transformer loads from models/diffusion_models safetensors (bf16 and fp8) into a ComfyUI ModelPatcher, so stock samplers, guiders and VAE nodes work with it — KSampler, BasicGuider/CFGGuider, Load VAE, plus FIBO2 nodes for conditioning (Qwen3-VL), sigmas and latents. The original diffusers-wrapper nodes below remain for the HF-repo workflow.

Native nodes (`fibo2_native/`, category FIBO2)

Node	What it does
FIBO2 Diffusion Model Loader	Loads any FIBO 2.0 checkpoint from `models/diffusion_models` into a ModelPatcher — the official Comfy-Org/Fibo-2.0 files (bf16 and int8_convrot, fused-key layout, dequantized on load), theunlikely's diffusers-layout bf16, and akhaliq/Fibo-2.0-edit-turbo-ComfyUI fp8_e4m3fn. Layout and quantization are detected automatically.
FIBO2 Text Encoder (Qwen3-VL)	Qwen3-VL-4B with the reference hidden-state bundling (6 × 3 layers, RMS-normalized), ported from BriaFibo2Pipeline.
FIBO2 Conditioning	Structured-JSON caption for generation, or edit instruction + reference images + mask (greys out masked regions as in training); encodes references through the VAE as context latents. Outputs positive + negative (trained null caption) conditioning.
FIBO2 Sigmas	The exact pipeline schedule: linspace(1, 1/steps) with FLUX-style mu shift from resolution.
FIBO2 Empty Latent	Packed 128-channel latents matching ComfyUI's Flux2 layout.

Wire them into a SamplerCustom flow (RandomNoise → FIBO2 Empty Latent, BasicGuider at CFG 1 for edit-turbo, Euler sampler, FIBO2 Sigmas) and decode with the stock FLUX.2 VAE.

Validated end-to-end: the native wrapper is bit-identical to the reference diffusers transformer (max abs diff 0.0), and this 4-step edit at 1024×1024 ran in ~1.2 s on an RTX PRO 6000 — from a flat gradient + circle to a stormy night, composition preserved. See tests/headless_test.py.

Diffusers-wrapper nodes (category FIBO)

Node	What it does
FIBO 2.0 Edit (local)	`Fibo-2.0-edit-turbo`: 4-step, guidance-free editing. Plain-text instructions (`make the sky stormy`) or VGL JSON with an `edit_instruction` key. Batched IMAGE input = multi-reference edit (`<image_1>` … in batch order, up to 4). Optional MASK for inpainting (single image only).
FIBO 2.0 Generate (local)	`Fibo-2.0-dev`: text-to-image from structured JSON prompts, with the model-card negative prompt built in.
FIBO Structured Prompt	Builds the structured JSON prompt from simple fields (subject, background, lighting, camera, style) so you never hand-write JSON. Wire its output into Generate.

Install

cd ComfyUI/custom_nodes
git clone https://huggingface.co/akhaliq/Fibo-2.0-ComfyUI
pip install -r Fibo-2.0-ComfyUI/requirements.txt
pip install -r Fibo-2.0-ComfyUI/fibo2_native/requirements.txt  # native nodes

Both models are gated: accept the licenses on briaai/Fibo-2.0-dev and briaai/Fibo-2.0-edit-turbo, then huggingface-cli login (or set HF_TOKEN) before starting ComfyUI. Weights (~16 GB per model) download on first use. Restart ComfyUI after installing.

An example workflow ships in example_workflows/fibo_edit.json — drag it onto the canvas.

Weights

Recommended: the official Comfy-Org/Fibo-2.0 files (fibo2_edit_turbo_bf16 or fibo2_edit_turbo_int8_convrot for ~half the download) into ComfyUI/models/diffusion_models/. Our fp8_e4m3fn is an alternative for Hopper GPUs. A drag-in workflow ships at fibo2_native/example_workflows/fibo2_native_edit.json.

VRAM

Both pipelines load with enable_model_cpu_offload(), streaming modules to the GPU on demand, so a ~12 GB card can run them. A lighter fp8 transformer (halves the download and, for runtimes that keep fp8 resident, the VRAM) is at akhaliq/Fibo-2.0-edit-turbo-fp8 — point the node's model_id widget at it, or set FIBO_EDIT_MODEL / FIBO_T2I_MODEL env vars.

Notes

BriaFibo2Pipeline is pinned to the verified diffusers PR #14991 head until it merges into a diffusers release; requirements.txt handles this.
FIBO models are CC-BY-NC 4.0 — non-commercial use, attribution to Bria. This node package is MIT-licensed code and contains no weights.
Model page: briaai/Fibo-2.0-edit-turbo · Paper: arXiv:2511.06876

akhaliq/Fibo-2.0-ComfyUI