H3 can also be used as an image edit model and its surprisingly good.
Prompting
Tell the model what you want to change / add / remove, then tell it what to keep.
Change the wave explosion in the background into an erupting volcano. Preserve the composition, lighting, style, facial expressions and backgrounds.
Resolution
Note that H3 only accepts specific image sizes natively, if your image is of a different size/resolution it is automatically rescaled but the result might have a difference in quality.
608 x 352
736 x 416
864 x 480
960 x 544
1056 x 608
1152 x 640
1216 x 672
1280 x 736
1344 x 768
1376 x 768
1504 x 832
1664 x 928
824 x 1024
1920 x 1088
Performance
At 1280p it takes me around 25 seconds, I'm using an RTX 5080 16gigs and 64gigs RAM.
Description
FAQ
Comments (9)
I'm going to check this out.
I've been experimenting with the ref model using the Reference to video node as an image editor. It almost feels like nano banana with repose and changing backgrounds and clothes. It maintains character consistency better than any of the true image edit models I've use locally.
The only bad thing I'm getting is that the output quality isn't there for me. It's blocky, but I see a lot of potential here. Someone smarter can figure it out.
Definitely a lot of potential, its only day 1 and so much has come out already.
In the meantime I'm playing around with SEGS and masks to see if anything cool can happen.
Edit; if you disconnect everything but the clip and vae inputs from the "Minimax H3 Image To Video" node, and set the frames to 5, it becomes a single image T2I generator. 35 seconds on my ten GB 3080 for 1024x1024
thanks for sharing. try swap the model to the ref2va model, i'm getting even better results
Agree, but I ask Claude to build a new workflow based on the default ref2va and this one. It looks like the best photo edit tool for "special" content. Just put hi-res photos as references, write a prompt in human language and generate 2k+ photo. Result is awesome, consistance depends on references quality and your prompt.
Oh I didn't know minimax could do image edits. Will try it right away!
It works amazingly well, new best video model AND image edit model? We are eating good this week!
You definitely do lose some quality no matter what you do in my experience but this is actually so cool regardless. Just randomly injecting spongebob into my images is super funny
It has some issues with stronger image edits request, but overall, this is a nice improvement for an image editor with the reference model.