ByteDance and Texas A&M research method distilling image and video diffusion models into few-step generators adversarially.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
DMAD (Distribution Matching as Adversarial Distillation) is a research method from ByteDance and Texas A&M University for distilling diffusion models into few-step student generators. It recasts distribution matching as a classification problem, using two discriminator heads on a shared backbone to compare teacher and student outputs.
The project page reports one-step image generation on ImageNet-64×64 with an FID of 1.24, or 1.04 with a projected discriminator variant, against 1.36 for the EDM teacher. It also reports four-step results for SDXL (FID 14.47 on COCO-10K), Wan2.1-T2V-14B (VBench 85.15) and a 33B MiniMax-H3 audio-video model. In a human evaluation on MiniMax-H3, the 4-step DMAD model was preferred over rCM in 84.6% of non-tie comparisons.
A YouTube creator reported that code was released with LoRAs of about 1.4 GB each. The provided material did not include a repository link, so this is unverified here. The work is a research release and not a hosted product. The page lists an arXiv paper (arXiv:2610.02188) but no direct link.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
A distillation method that lets video models like Minimax generate in four steps instead of 20-30, using discriminator heads to compare teacher and student outputs. Code released with LoRAs of 1.4 GB each.
Closest neighbours by meaning, not just by category.
Research · ByteDance
ByteDance research method for few-step diffusion distillation, adding a one-line projection to the DMD update.
Research · ByteDance
ByteDance research method for post-training image diffusion models using on-policy self-distillation from reward signals.
Research · ByteDance
ByteDance research framework that distills multiple specialist image models into one via on-policy velocity matching.
Model
One- and two-step image generation model distilled from Z-Image-Turbo using text-stream adversarial distillation.
Model · ByteDance
ByteDance's open-weights one-step diffusion transformer for restoring and upscaling video and images.
Model · ByteDance
ByteDance unified video generation and editing framework pairing an MLLM semantic planner with a DiT renderer.
