Test, comparison with the original MiniMax H3
Thanks for the model, it's really impressive!
Here's a quick comparison with the original MiniMax H3.
The level of detail has improved significantly, and it now feels like a fully capable image generation model. This is what we were missing.
The only thing I've noticed is that character likeness seems slightly worse. The video model still does a better job of preserving character identity and resemblance.
Which is a bit of a shame, since that was one of the model's strong points.
Yeah, the video model was able to make character sheets almost perfectly with references and grabbing the first frame. This image model is faster even with higher resolution if you know your prompt and skip the enhancer, but the face likeness is not good most of the time. At least the couple of video LoRAs I tried did work for the image model, so I love that compatibility.
What I notice is a weird noise in the h3-image generation that is not present in the H3 video model, is there a way to avoid it?
It breaks the generation and makes the image not properly usable.
How about H3-image vs H3 on video gen?
What I notice is a weird noise in the h3-image generation that is not present in the H3 video model, is there a way to avoid it?
It breaks the generation and makes the image not properly usable.
Could the authors fix VAE or model so there are no weird arfifacts?
Edit mode has nasty artifacts if you are on 5+ step - euler or lcm, does not matter. It always adds an ugly patterm to the whole image. It is clearly seen in blank areas like grey background or when you generate a few images with the same seed and only change something - the destructive patter across the whole image is the same.
There is a certain situation in workflow when this destruction does not happen. I am not sure what causes that background pattern which sinks into the whole image.
Could be something with resolutions or format of reference.