Please upload the original model, not the “Pruned” version.
Please upload the original model, not the “Pruned” version.
In my tests, Minimax H3 for video produces the best results with the unpruned model. This will likely be the case with the Image model as well. I'd like to be able to choose which model. Thank you for your work.
That's not actually possible, because the way the model is pruned is functionally identical.
I also thought this for a second when the video model came out, but it turns out the cause of this was implementation issues present at the time.
Even MiniMax themselves say that the AdaLN-related branches aren't necessary for inference, because they can be precomputed and cached; ComfyUI actually went further with the LUT curve.
If you insist, I'm sure it wouldn't be too much trouble to get on Marimo or something and have their AI agent make a Comfy-format version of the unpruned weights, but I guarantee you any amount of quantization is far more destructive.
The Minimax developers have released the full model weights, and I think it would be right if we could run them natively, just as we always have in the past.