Fibo 2.0 T8

Fibo 2.0 T8 for ComfyUI

English · 简体中文

Run Fibo 2.0 Dev text-to-image and editing, and Fibo 2.0 Edit Turbo editing, through BRIA's official Diffusers pipeline. DiT, VAE and Qwen3-VL text encoder have separate files and separate loader nodes.

• Supports prepared BRIA checkpoints and Comfy-Org single files, including BF16 and INT8 ConvRot.
• Plain text or structured JSON prompts; 1–5 editing references; single-reference masks; automatic or supplied text images.
• CPU offload, sequential generation, seed and execution metadata. No external prompt API is required.

Installation

In ComfyUI-Manager, search for Fibo 2.0 T8 (publisher t8star, node ID fibo-v2-t8), install, then restart ComfyUI. Registry publication follows the official ComfyUI process.

Manual installation:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-Fibo-V2-t8.git
python -m pip install -r Comfyui-Fibo-V2-t8/requirements.txt

Use ComfyUI's Python for the pip command; Windows portable builds use python_embeded/python.exe. Requires Python 3.11+, a CUDA PyTorch installation and a compatible NVIDIA GPU. BF16 with model_cpu offload is the recommended starting point. Dependencies include a pinned Diffusers commit with BriaFibo2Pipeline; an older stable Diffusers release will not work.

Models

Download from Comfy-Org/Fibo-2.0 and keep the published filenames:

| ComfyUI folder | File(s) | |---|---| | models/diffusion_models/ | fibo2_dev_bf16.safetensors or fibo2_dev_int8_convrot.safetensors | | models/diffusion_models/ | fibo2_edit_turbo_bf16.safetensors or fibo2_edit_turbo_int8_convrot.safetensors | | models/text_encoders/ | qwen3vl_4b_bf16.safetensors or qwen3vl_4b_int8_convrot.safetensors | | models/vae/ | flux2-vae.safetensors |

Only download the DiT variant(s) you need. TE and VAE are shared. Extra model paths and subfolders are supported.

First run: the plugin verifies the source SHA256 and creates separate official-layout component caches under models/fibo2/converted/. BF16 adaptation preserves tensor values. INT8 ConvRot is dequantized to BF16 for the official pipeline: it retains quantization error and uses BF16-sized RAM/VRAM, rather than native ComfyUI INT8 kernels. Allow approximately 16.3 GB per DiT, 8.9 GB for TE and 0.34 GB for VAE in additional disk cache. Support assets are bundled; execution needs no additional model download. Sources are kept intact.

For original BRIA snapshots and existing prepared assets, see advanced setup.

Workflows

Drag a JSON from examples/workflows into ComfyUI. Select a DiT, VAE and TE in the three loaders, connect them to Fibo 2.0 Assemble Official Pipeline, then run a generation or editing node. Copy example inputs into ComfyUI/input for editing examples. API graphs are also included.

| Task | Node | Model-card starting recipe | |---|---|---| | Text-to-image | Fibo 2.0 Dev Generate | Dev · 50 steps · CFG 5 | | Image editing | Fibo 2.0 Dev Edit | Dev · 50 steps · CFG 5 | | Fast image editing | Fibo 2.0 Turbo Edit | Edit Turbo · 4 steps · CFG 1 |

auto uses trained resolution around 1 MP. Explicit sizes must be multiples of 16. Masks guide editing; they do not guarantee unchanged pixels outside the mask. These pipeline nodes use their own component types; outputs are standard ComfyUI IMAGE tensors.

Links

Bilibili · YouTube · API · Free gallery · Online AI apps · ComfyUI bundle · Hugging Face

License and credits

Plugin code: Apache-2.0. Model weights and bundled support assets retain their upstream licenses; Fibo model cards specify CC BY-NC 4.0 with commercial licensing through BRIA. See third-party notices. Thanks to BRIA, Comfy-Org and Hugging Face Diffusers.