Home

Awesome

Extra Models for ComfyUI

This repository aims to add support for various different image diffusion models to ComfyUI.

Installation

Simply clone this repo to your custom_nodes folder using the following command:

git clone https://github.com/city96/ComfyUI_ExtraModels custom_nodes/ComfyUI_ExtraModels

You will also have to install the requirements from the provided file by running pip install -r requirements.txt inside your VENV/conda env. If you downloaded the standalone version of ComfyUI, then follow the steps below.

Standalone ComfyUI

I haven't tested this completely, so if you know what you're doing, use the regular venv/git clone install option when installing ComfyUI.

Go to the where you unpacked ComfyUI_windows_portable to (where your run_nvidia_gpu.bat file is) and open a command line window. Press CTRL+SHIFT+Right click in an empty space and click "Open PowerShell window here".

Clone the repository to your custom nodes folder, assuming haven't installed in through the manager.

git clone https://github.com/city96/ComfyUI_ExtraModels .\ComfyUI\custom_nodes\ComfyUI_ExtraModels

To install the requirements on windows, run these commands in the same window:

.\python_embeded\python.exe -s -m pip install -r .\ComfyUI\custom_nodes\ComfyUI_ExtraModels\requirements.txt

To update, open the command line window like before and run the following commands:

cd .\ComfyUI\custom_nodes\ComfyUI_ExtraModels\
git pull

Alternatively, use the manager, assuming it has an update function.

Sana

Original Repo

[!CAUTION] As many people have had issues with Sana, it's for now recommended to try the fork by the Sana devs, which auto downloads all models:

Readme | Fork repo

A full rewrite to have better integration is in progress in this PR but isn't ready yet.

https://github.com/NVlabs/Sana/blob/main/asset/docs/ComfyUI/comfyui.md https://github.com/Efficient-Large-Model/ComfyUI_ExtraModels

Model info / implementation

Usage

  1. Download the model weights from the Sana HF repo - the HF account has alternative models available too.
  2. Place them in your checkpoints folder
  3. Load them with the correct PixArt checkpoint loader
  4. Use the "Gemma Loader" node - it should automatically download the requested model from Huggingface - Recommended to use the 4bit quantized model on CPU when low on memory.
  5. Download the VAE from here or here and place it in your VAE folder after renaming it.
  6. Use either the "Empty Sana Latent Image" or "Empty DCAE Latent Image" node for the latent input when doing txt2img.

Sample workflow

Sana

PixArt

Original Repo

Model info / implementation

Usage

  1. Download the model weights from the PixArt alpha repo - you most likely want the 1024px one - PixArt-XL-2-1024-MS.pth
  2. Place them in your checkpoints folder
  3. Load them with the correct PixArt checkpoint loader
  4. Follow the T5v11 section of this readme to set up the T5 text encoder

[!TIP] You should be able to use the model with the default KSampler if you're on the latest version of the node. In theory, this should allow you to use longer prompts as well as things like doing img2img.

Limitations:

[!IMPORTANT]
Installing xformers is optional but strongly recommended as torch SDP is only partially implemented, if that.

Sample workflow here

PixArtT12

PixArt Sigma

The Sigma models work just like the normal ones. Out of the released checkpoints, the 512, 1024 and 2K one are supported.

You can find the 1024 checkpoint here. Place it in your models folder and select the appropriate type in the model loader / resolution selection node.

[!IMPORTANT] Make sure to select an SDXL VAE for PixArt Sigma!

PixArt LCM

The LCM model also works if you're on the latest version. To use it:

  1. Download the PixArt LCM model and place it in your checkpoints folder.
  2. Add a ModelSamplingDiscrete node and set "sampling" to "lcm"
  3. Adjust the KSampler settings - Set the sampler to "lcm". Your CFG should be fairly low (1.1-1.5), your steps should be around 5.

Everything else can be the same the same as in the example above.

PixArtLCM

HunYuan DiT

WIP implementation of HunYuan DiT by Tencent

The initial work on this was done by chaojie in this PR.

Instructions:

You may also try the following alternate model files for faster loading speed/smaller file size:

You can use the "simple" text encode node to only use one prompt, or you can use the regular one to pass different text to CLIP/T5.

Sample Workflow

image

DiT

Original Repo

Model info / implementation

Usage

  1. Download the original model weights from the DiT Repo or the converted FP16 safetensor ones from Huggingface.
  2. Place them in your checkpoints folder. (You may need to move them if you had them in ComfyUI\models\dit before)
  3. Load the model and select the class labels as shown in the image below
  4. Make sure to use the Empty label conditioning for the Negative input of the KSampler!

ConditioningCombine nodes should work for combining multiple labels. The area ones don't since the model currently can't handle dynamic input dimensions.

Sample workflow here

DIT_WORKFLOW_IMG

T5

T5v11

The model files can be downloaded from the DeepFloyd/t5-v1_1-xxl repository.

You will need to download the following 4 files:

Place them in your ComfyUI/models/t5 folder. You can put them in a subfolder called "t5-v1.1-xxl" though it doesn't matter. There are int8 safetensor files in the other DeepFloyd repo, thought they didn't work for me.

For faster loading/smaller file sizes, you may pick one of the following alternative downloads:

To move T5 to a different drive/folder, do the same as you would when moving checkpoints, but add t5: t5 to extra_model_paths.yaml and create a directory called "t5" in the alternate path specified in the base_path variable.

Usage

Loaded onto the CPU, it'll use about 22GBs of system RAM. Depending on which weights you use, it might use slightly more during loading.

If you have a second GPU, selecting "cuda:1" as the device will allow you to use it for T5, freeing at least some VRAM/System RAM. Using FP16 as the dtype is recommended.

Loaded in bnb4bit mode, it only takes around 6GB VRAM, making it work with 12GB cards. The only drawback is that it'll constantly stay in VRAM since BitsAndBytes doesn't allow moving the weights to the system RAM temporarily. Switching to a different workflow should still release the VRAM as expected. Pascal cards (1080ti, P40) seem to struggle with 4bit. Select "cpu" if you encounter issues.

On windows, you may need a newer version of bitsandbytes for 4bit. Try python -m pip install bitsandbytes

[!IMPORTANT]
You may also need to upgrade transformers and install spiece for the tokenizer. pip install -r requirements.txt

MiaoBi

Original from:

Instructions

这是妙笔的测试版本。妙笔，一个中文文生图模型，与经典的stable-diffusion 1.5版本拥有一致的结构，兼容现有的lora，controlnet，T2I-Adapter等主流插件及其权重。

This is the beta version of MiaoBi, a chinese text-to-image model, following the classical structure of sd-v1.5, compatible with existing mainstream plugins such as Lora, Controlnet, T2I Adapter, etc.

Example Prompts:

Example Workflow

Example Workflow (diffusers)

MiaoBi

VAE

A few custom VAE models are supported. The option to select a different dtype when loading is also possible, which can be useful for testing/comparisons. You can load the models listed below using the "ExtraVAELoader" node.

Models like PixArt/DiT do NOT need a special VAE. Unless mentioned, use one of the following as you would with any other model:

Consistency Decoder

Original Repo

This now works thanks to the work of @mrsteyk and @madebyollin - Gist with more info.

Deflickering Decoder / VideoDecoder

This is the VAE that comes baked into the Stable Video Diffusion model.

It doesn't seem particularly good as a normal VAE (color issues, pretty bad with finer details).

Still for completeness sake the code to run it is mostly implemented. To obtain the weights just extract them from the sdv model:

from safetensors.torch import load_file, save_file

pf = "first_stage_model." # Key prefix
sd = load_file("svd_xt.safetensors")
vae = {k.replace(pf, ''):v for k,v in sd.items() if k.startswith(pf)}
save_file(vae, "svd_xt_vae.safetensors")

AutoencoderKL / VQModel

kl-f4/8/16/32 from the compvis/latent diffusion repo.

vq-f4/8/16 from the taming transformers repo, weights for both vq and kl models available here

vq-f8 can accepts latents from the SD unet but just like xl with v1 latents, output largely garbage. The rest are completely useless without a matching UNET that uses the correct channel count.

VAE_TEST