Skip to content
Trending storyDeveloping

ModelScope open-sources DMAD few-step distillation for image, video and audio-video generation

1 article1 sourcesince Oct 12Last article 2h ago ·

Overview

AISummary of 1 article

ModelScope says it has open-sourced DMAD, a distillation method that cuts image, video and joint audio-video generation to 1 to 4 steps.

In its X post, ModelScope says a 1.4 GB rank-128 LoRA turns the 50-step MiniMax-H3 into a 4-step generator of 1344x768 video with stereo audio.

The release includes weights for MiniMax-H3, Wan2.1, SDXL and ImageNet-64. ModelScope says 15-second clips run on a 16 GB GPU and 5-second clips on 8 GB, with 64 GB of system RAM recommended. The weights inherit their base model licenses, and the MiniMax-H3 LoRAs fall under the MiniMax H3 Community License.

Written by AI from the articles below · updated Oct 12, 9:22 AM ET

Check the sources:

Latest developmentOct 12, 9:12 AM ET
ModelScope open-sources DMAD, a 1 to 4 step distillation method for image, video, and audio-video generation

Article timeline

The articles in this story. Times are ET.

Oct 12
  1. ModelScopeOfficial
    ModelScope open-sources DMAD, a 1 to 4 step distillation method for image, video, and audio-video generation

    AIModelScope says DMAD is now open source and enables 1 to 4 step generation for image, video, and joint audio-video models. The post says a 1.4 GB rank-128 LoRA turns the 50-step MiniMax-H3 into a 4-step generator of 1344x768 video with stereo audio. It says 15-second clips run on a 16 GB GPU and 5-second clips on 8 GB, with 64 GB of system RAM recommended. Weights for MiniMax-H3, Wan2.1, SDXL, and ImageNet-64 are released, and the MiniMax-H3 LoRAs fall under the MiniMax H3 Community License.

    GIF from @ModelScope2022's post

Heat trend

Not enough continuous observations to show a trend yet.