Skip to content

TencentQQGYLab/ComfyUI-ELLA

Repository files navigation

ComfyUI-ELLA


ComfyUI implementation for ELLA.

🌟 Changelog

  • [2024.4.30] Add a new node ELLA Text Encode to automatically concat ella and clip condition.
  • [2024.4.24] Upgraded ELLA Apply method. Better compatibility with the comfyui ecosystem. Refer to the method mentioned in ComfyUI_ELLA PR #25
    • DEPRECATED: Apply ELLA without simgas is deprecated and it will be removed in a future version.
  • [2024.4.22] Fix unstable quality of image while multi-batch. Add CLIP concat (support lora trigger words now).
  • [2024.4.19] Documenting nodes.
  • [2024.4.19] Initial repo.

📌 Notice

  • SIGMAS from node BasicScheduler or TIMESTEPS by node Set ELLA Timesteps must be the same as the KSampler settings. Because Timestep-Aware Semantic Connector (TSC), which dynamically adapts semantics features over sampling time steps, has been introduced.
  • If you need concat clip CONDITIONING to make LoRA trigger words effective, ELLA output CONDITIONING always needs to be linked to the conditioning_to of Conditioning (Concat) node.

📚 Example workflows

The examples directory has workflow examples. You can directly load these images as workflow into ComfyUI for use.

workflow_example

All legacy workflows was compatible. But it is deprecated and will be removed in a future version.

workflow_example_legacy

🎉 It works with controlnet!

workflow_controlnet

🎉 It works with lora trigger words by concat CLIP CONDITIONING!

⚠️ NOTE again that ELLA CONDITIONING always needs to be linked to the conditioning_to of Conditioning (Concat) node.

workflow_lora

With ELLA Text Encode node, can simplify the workflow.

With the upgrade(2024.4.24), some interesting workflow can be implemented, such as using ELLA only in positive. As shown below:

workflow_lora_positive_ella_only

positive + negative positive only
ella_with_lora ella_with_lora_positive_only

However, there is no guarantee that positive-only will bring better results.

Workflow with AYS.

workflow_ella_ays

AYS got with more visual details and better text-alignment, ref to paper.

w/ AYS w/o AYS

And EMMA is working in progress.

📗 Install

Download or git clone this repository inside ComfyUI/custom_nodes/ directory. ComfyUI-ELLA requires the latest version of ComfyUI. If something doesn't work be sure to upgrade.

cd ComfyUI/custom_nodes
git clone https://github.com/TencentQQGYLab/ComfyUI-ELLA

Next install dependencies.

cd ComfyUI-ELLA
pip install -r requirements.txt

📙 Models

These models must be placed in the corresponding directories under models.

Remember you can also use any custom location setting an ella & ella_encoder entry in the extra_model_paths.yaml file.

  • ComfyUI/models/ella, create it if not present.
  • ComfyUI/models/ella_encoder, create it if not present.

In summary, you should have the following model directory structure:

ComfyUI/models/ella/
└── ella-sd1.5-tsc-t5xl.safetensors

ComfyUI/models/ella_encoder/
└── models--google--flan-t5-xl--text_encoder
    ├── config.json
    ├── model.safetensors
    ├── special_tokens_map.json
    ├── spiece.model
    ├── tokenizer_config.json
    └── tokenizer.json

📖 Nodes reference

Nodes reference

🔍 Common promblem

  • XXX not implemented for 'Half'. See issue #12
  • AYS + Ella getting dark image generations. See issue #39
    • Check if add_noise of SamplerCustom node is enabled.
    • Lower the cfg of SamplerCustom node.

📝 TODO

  • Support prompt weighting

🤗 Contributors (direct & indirect)


JettHu

budui

kijai

huagetai

😋 Thanks

😉 Citation

@misc{hu2024ella,
      title={ELLA: Equip Diffusion Models with LLM for Enhanced Semantic Alignment}, 
      author={Xiwei Hu and Rui Wang and Yixiao Fang and Bin Fu and Pei Cheng and Gang Yu},
      year={2024},
      eprint={2403.05135},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}