AI

DMAD: Distribution Matching as Adversarial Distillation for Fast Visual Generation

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 20:18:24 UTC.

Briefings

DMAD improves visual generation by replacing auxiliary score fitting with a classification-based discriminator that learns log-density ratios directly. It achieves a FID of 1.04 on ImageNet-64x64 and a VBench score of 85.15 on Wan2.1-T2V-14B, outperforming existing few-step methods. The method uses a shared backbone with two discriminator heads to distinguish real data from student samples, optimizing via linear losses on logits.

Key points

• DMAD introduces a classification-based discriminator head that learns log-density ratios directly, replacing the auxiliary score fitting required by standard Distribution Matching Distillation.
• The method achieves a FID of 1.04 on ImageNet-64x64 and a VBench total score of 85.15 on Wan2.1-T2V-14B, representing the best few-step performance among compared methods.

Context

DMAD addresses the high computational cost of auxiliary diffusion models in distribution matching by using a shared backbone with two discriminator heads. It recasts distribution matching as classification, learning log-density ratios directly through linear losses on logits, with gap-based reweighting adapting teacher supervision across noise levels.

Limitations

Results are limited to ImageNet-64x64 and COCO-10K benchmarks; performance on other datasets or video generation tasks is not reported. The study compares few-step methods against multi-step teachers but does not provide a full multi-step baseline for comparison.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/11/2026, 18:10:34 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeifx4ib5vatwrs3l5rmsnvczqevkqad6ksk7yksrej62bnb7jtrvyy