LongCat Image & Image Edit in ComfyUI (Patreon)

Video - https://youtu.be/0gvmND1mJrE

Overview

ComfyUI now has native support for the 6B LongCat Image and LongCat Image Edit models!

At just 12.5GB, it is a much smaller download than the 40GB filesize for similar Image Edit models such as Qwen Image Edit 2511. If you struggled to use Qwen Image Edit even at FP8, then you may wish to try LongCat.

Note also that if you've tried via the Comfy Example Templates and thought LongCat was a bit pants, then chances are you'll want to give it another go because the example workflows don't really let you get the best out of the model! My first attempts with the examples came out pretty blurry, but that is soon fixed with a bit of masking - plus the latent encoding method means you aren't actually limited to just 1 megapixel.

Workflow Information

To use this workflow you will still need to install a few custom nodes, as detailed in the in-workflow note:

Model download links are also included:

For the LongCat Image Edit Turbo model, they have provided a Convert Script to change it from Diffusers format into Comfy format.

You can download the original LongCat Image Edit model from huggingface:

https://huggingface.co/meituan-longcat/LongCat-Image-Edit-Turbo/resolve/main/transformer/diffusion_pytorch_model.safetensors?download=true

Run the script to convert the model and you'll be able to enjoy the 8-step goodness.

Personally, I still prefer the non-Turbo model as there are some image editing tasks Turbo doesn't do quite as well on, but the choice is yours!

I've also included some bookmarks, so if you press 0 then you'll go straight to the model loading group:

You can use either the Image or Image Edit models all via the single workflow.

Bookmark 1 will take you to the settings:

Pick one of the 4 latent options available:

The Empty Latent is great from when you don't care about pixel-perfect changes and are instead looking to do things like change aspect ratios or image style. You can select the Empty Latent Size below:

The Image latent is the default for ComfyUI workflows and will VAE encode the image, using the image size for your generation.

When using any of the image latents, there are a number of encoding options:

If you're using the Prompt Text Encoding method, then you should set your image size to 1 megapixel (e.g. 1280x720). Any larger, and the editing task may not be the best.

If you're using the Latent Encoding method, then other resolution are unlocked, up to around 1.5 megapixels is best, but the model can just about handle sizes up to 2048x2048.

The Background Removed option will automatically mask the background which is ideal if you're just changing that.

The Masked Image option will use the image mask you prompt for or provide. Bookmark 2 will take you to image and prompt selection:

There is a switch depending on if you want to use the prompt-generated automask, or your own manual masking:

The preview image will show you the selected mask.

Using a mask is highly recommended if you want specific things in an image (such as hair colour) to change, but you don't want other things in the image to change.

LongCat Image likes long, detailed prompts so you may wish to use an LLM of your choice for this:

Ollama can be installed via https://ollama.com and there are LLMs available in many different sizes.

You may also wish to enable Image Compare or Image Upscaling Options:

If you are comparing images, don't forget to set which 2 images you'd like to compare via the GET nodes:

Examples:

Image Edit, Empty Latent, 1280x1280, Latent Encoding, "change the image style to an impasto, oil painted style", 30 steps, CFG 1.5:

Prompt Encoding, "give her long, curly blonde hair", 1088x1280, CFG 4:

Latent Encode, masked image, "change the woman's face into a rodent's face":

Enjoy!