Differences Between Pruned and Unpruned?
Thanks for the hardwork,
AI says: Pruned means some of the model's weights/parameters have been removed entirely — not just compressed or quantized to a smaller data type, but deleted from the network.
Removing some parts of the full model lets it run on machines with less VRAM. I remember SD1.5 used this method.
https://huggingface.co/MiniMaxAI/MiniMax-H3
For scalability and generalization, we adopt a relatively simple Transformer block design. H3-Omni-Transformer is a 33B-parameter dense, single-stream Transformer, with approximately 13B parameters residing in AdaLN-related branches. Because the AdaLN modulation outputs can be precomputed and cached, these parameters do not need to be loaded for inference-only deployment. We release the complete model weights to support further development, including fine-tuning.
We found that the model's modulation weights (~40% of the total parameters) could be pruned and replaced with a functionally equivalent lookup table, dramatically shrinking the memory footprint with no loss in output quality. See blog post: https://blog.comfy.org/p/minimax-h3-day-0-support-in-comfyui
Yes it's indeed without the modulation weights, it remains functional since we basically compressed that into a curve derived from the weights, so even the precomputation isn't necessary. It does change the outputs of the model on same seed though, but based on my testing not really to the worse or better, just different. Some examples:
Yes it's indeed without the modulation weights, it remains functional since we basically compressed that into a curve derived from the weights, so even the precomputation isn't necessary. It does change the outputs of the model on same seed though, but based on my testing not really to the worse or better, just different. Some examples:
@kijai kijaiii what resolution did you generated these videos?