ComfyUI Extension: ComfyUI_RH_UNO
This is a UNO ComfyUI plugin implementation that can run the full version with 24GB VRAM, as well as quickly run the FP8 version.
Custom Nodes (0)
README
中文版说明.
UNO (Unity and Novel Output) for ComfyUI
This repository hosts the ComfyUI implementation of UNO (Unity and Novel Output), supporting FLUX models. This implementation includes several new features and optimizations. That can run the full version with 24GB VRAM, as well as quickly run the FP8 version.
Online Access
You can also access RunningHub online to use this plugin and model for free. Run&Download this Workflow: https://www.runninghub.ai/post/1910316871583789058
Features
- Updated to match author's latest version
- Support for
flux-dev-fp8
andflux-schnell-fp8
- Note:
flux-schnell-fp8
offers lower consistency but much faster generation (4 steps)
- Support for
- Memory optimization through block swapping
- Run BF16 models on 24GB GPUs
- Support for both
flux-dev
andflux-schnell
in BF16 mode on 24GB GPUs
- Progress bar to display denoising progress in real-time
- Local model loading configured via
config.json
Model Configuration
Directory Structure
Models are configured in the root config.json
file. The default structure expected is:
ComfyUI/models
flux/
FLUX.1-schnell/ ###download from https://huggingface.co/black-forest-labs/FLUX.1-schnell
text_encoder/
tokenizer/
text_encoder_2/
tokenizer_2/
unet/
flux1-schnell.sft
flux1-dev.sft
vae/
ae.safetensors
UNO/
dit_lora.safetensors
Text Encoder and CLIP Configuration
For T5 and CLIP models, there are two organization options:
-
Single directory (XLabs-AI/xflux_text_encoders style):
- Set
"t5-in-one": 1
or"clip-in-one": 1
- Text encoder and tokenizer are in the same folder
- Set
-
Official structure (separate directories):
- Set
"t5-in-one": 0
or"clip-in-one": 0
- Point
"t5"
or"clip"
to the parent directory - Child directories must follow official naming conventions:
- CLIP:
text_encoder
andtokenizer
- T5:
text_encoder_2
andtokenizer_2
- CLIP:
- Set
Other Model Paths
-
VAE:
- Default:
comfyui/models/vae/
- Configure with
"vae_base"
in config.json
- Default:
-
FLUX Models:
- Default:
comfyui/models/unet/
- Configure with
"model_base"
in config.json - Note: The author believes current FP8 FLUX models have issues, so BF16 models are used in both modes
- Default:
-
DIT-LoRA Models:
- Default:
comfyui/models/UNO/
- Configure with
"lora_base"
in config.json
- Default:
Example
Acknowledgments
Thanks to the original author. Visit the official repository at: https://github.com/bytedance/UNO