Mamad8

/

MiniMax-H3-Image-VAE

like 62

Pretty much any ref2va workflow is fine. You just need to alter "\comfy_extras\nodes_minimax_h3.py" to accept 1 as a frame number (the other discussions on this model show you how to do that. It's just copy and paste). Then you disconnect any math node that's connected to the "length" slot of the ref2v node so that you can set it to 1 frame. Then set the output to PNG. Then swap the default VAE to this one.

The screen shot below is an example. It left the initial image intact and only made the changes that I requested in the prompt.

The prompt I used was simply:

Subjects And References:

<subject 1> is the lumberjack man in <Picture 1>.
<subject 2> is the blue dress in <Picture 2>.

Tasks:

Edit <Picture 1> so that <subject 1> is wearing <subject 2>.

In the screenshot you might notice that I'm using the "minimax_h3_hybrid_fl2va_ref2va_b20-49.safetensors" model. You can find it here: https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/tree/main . I'd HIGHLY recommend it for using MMH3 as an image generator/editor. I'm also using a LoRA called "minimax_h3_ref_lora_rank_256" to help with image editing tasks. You can find it here: https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras .

@fizzlepoof , could you share your workflow? Just upload it to HF. It would've been faster than typing all that ☺️

@positiveelevation lol I get what you're saying but it took me like 2 minutes to type that. Whereas uploading a workflow json without expanding the discussion page to Narnia requires a bit more effort on my end. I'm quite lazy.

I just tossed it in a dataset page here: https://huggingface.co/datasets/fizzlepoof/MMH3_Image_Edit_Workflow/tree/main .

This isn't my actual workflow since that one uses some unpublished custom nodes and is... large. This is just a basic example of a use case for this VAE.

Edit: I also added instructions to the readme in the dataset folder.

@fizzlepoof Thanks, it's working now. Made with your own workflow:

@fizzlepoof , it's weird, when I made the above picture, I took the meme of the cartoon dog and successfully prompted it to replace with a real human knight, and it worked.

Now I'm trying to prompt it to bring a comic book panel to realistic movie version, and failing miserably. What do you make of this?

You might need a more robust workflow to handle a "anything2real" style edit without adding a bunch of extra references, since the one I posted is so stripped down.

There's a chain of reddit posts that has a more complete workflow that doesn't use unpublished nodes: https://www.reddit.com/r/StableDiffusion/comments/1vo1ab3/h3_as_a_singleimage_edit_model/

I've haven't tried it but he is using the same VAE and MinMax model. He's also using a turbo LoRA though instead of Spectrum and he isn't using a detailer to increase the quality so I'm not certain how that works out but the examples look decent.

I tried his workflow and it's way worse than yours. I even tried bypassing the Turbo LoRA and setting steps to 40.
The only flaw in yours is style change. You could see from the chicken prompt that it does great edits.

Here's the chicken prompt and the actor prompt on the latest workflow from the post you linked, with latest Comfy I just updated:

I think I should just wait for the official MiniMax image model, in the meantime I can use your wf for simple edits, I found it better than Klein 9B and Qwen Image Edit.