ComfyUI-DreamO

DreamO Comfyui

DreamO ComfyUI native implementation.

arXiv demo

[!Important]
2025.06.24 - We are excited to release DreamO v1.1 with significant improvements in image quality, reduced likelihood of body composition errors, and enhanced aesthetics. Learn more about this model

Install

This implementation is based on the 2025.5.19 version of ComfyUI (commit ID: e930a38). Compatibility issues may occur if you're using an older version.

# manual install
cd custom_nodes
git clone https://github.com/ToTheBeginning/ComfyUI-DreamO.git
# Please make sure that you have installed the forked version of facexlib, not the original facexlib. Otherwise, you may encounter face parsing errors.
pip install -r requirements.txt
# restart comfyui

Models

FLUX models

If your machine already has FLUX models downloaded, you can skip this.

• Original bf16 model: dit, t5
• 8 bit FP8: dit, t5
• Clip and VAE (for all models): clip, vae

DreamO models

• Download all files ending with .safetensors in https://huggingface.co/ByteDance/DreamO/tree/main/comfyui => ComfyUI/models/loras.
• 🔥🔥v1.1: Download all files ending with .safetensors in https://huggingface.co/ByteDance/DreamO/tree/main/v1.1 => ComfyUI/models/loras.
• (Support auto-download) Download dreamo-embedding => ComfyUI/models/dreamo
• (Support auto-download) Download ben2 => ComfyUI/models/dreamo
• Download flux-turbo => ComfyUI/models/loras, and rename it to flux-turbo.safetensors

Workflows

Please fine the workflows in workflows folder, this v1.1 workflow is the latest version.

The v1.1 model, when combined with the Super-Realism LoRA, can further enhance realism. However, since it's a realism-focused LoRA, it may affect style transfer. Use it as needed.

Nodes

• DreamOProcessorLoader
  • This node loads two image preprocessing models: the BEN2 model for background removal and the facexlib model for aligned face detection.
• DreamORefEncode
  • This node encodes the reference image into a latent representation based on the selected task type. Three task types are available: ip, id, and style.
    • ip: will remove the backgound of the reference image
    • id: will align&crop the face from the reference image, similar to PuLID
    • style: will keep the backgound of the reference image. you still need trigger meta prompt to activate the style transfer task
• ApplyDreamO
  • This node adds a hook to the Flux model to support concatenating the reference latent with the noisy latent.

Note

• The current code does not implement the logic for true CFG, which means you need to set cfg=1 in the sampler node
• As mentioned