DMAD Cuts Visual Generation to Few Steps

// 54m agoRESEARCH PAPER

DMAD Cuts Visual Generation to Few Steps

DMAD reframes distribution matching as adversarial classification, eliminating DMD’s separately fitted auxiliary score model. The research reports strong one-step image, four-step image, video, and audio-video generation results at substantially lower inference cost.

// ANALYSIS

DMAD’s strongest contribution is a cleaner distillation objective that turns classifier log-density ratios into a direct training signal for fast visual generators.

–Reports 1.04 FID for one-step ImageNet-64 generation and 14.47 FID for four-step SDXL on COCO-10K
–Reaches 85.15 on VBench with four-step Wan2.1-T2V-14B generation
–Four-step MiniMax-H3 students achieve 79.1% preference over DMD2 and 84.6% over rCM for joint audio-video generation
–The speed-quality tradeoff remains workload-dependent: one-step image models lose local detail, while competing methods retain advantages on some semantic video metrics
–Open code and models make DMAD immediately relevant for researchers deploying low-latency image and video generation

// TAGS

dmaddistillationinferenceimage-genvideo-genmultimodalresearch

DISCOVERED

54m ago

2026-10-04

PUBLISHED

1h ago

2026-10-04

RELEVANCE

9/ 10

AUTHOR

AI Search

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL1h ago

Whistle Packs Multilingual Speech-to-Text into 16.9 MB.

Cactus Compute released Whistle, an open CPU speech-recognition model that transcribes seven languages, supports 30-second inputs, word timestamps, speech embeddings, and on-device processing. It shares the Needle runtime and weighs just 16.9 MB. [Cactus launch post](https://www.cactuscompute.com/blog/whistle)

RESEARCH1h ago

TactileStep Makes Humanoid Landings Softer

Tsinghua’s TactileStep adds sole-pressure feedback to Unitree G1 locomotion policies, using contact force, center of pressure, and loaded-area signals to regulate touchdown and stance. Tests across varied terrain report up to 48.8% lower touchdown force and 23.8% greater stance contact area. [Official project page](https://tactilestep.github.io/) · [Paper](https://arxiv.org/abs/2609.28959)

MODEL1h ago

Phonon-2 Releases 164MB Open ASR

Fermion Research released Phonon-2, an open English speech-recognition model with a 164 MB download, 5.21% average word error rate, and 174× realtime inference on an M5 MacBook Air. It runs locally across macOS, Linux, Windows, and NVIDIA GPUs. [Announcement](https://www.fermionresearch.com/research/phonon-2/)