ByteDance research method for few-step diffusion distillation, adding a one-line projection to the DMD update.
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.
PDMD (Projected Distribution Matching Distillation) is a diffusion step-distillation method from researchers at ByteDance and UC San Diego. It modifies the Distribution Matching Distillation (DMD) update with a one-line projection that filters the student's training error. According to the project page, nothing else changes: no auxiliary loss, no additional network and no extra model pass.
The project page reports results on video and audio generation. It shows MiniMax-H3-33B distilled to four network evaluations (and also to two), producing 1344×768 video at 24 fps with 345 frames. On a 387-prompt benchmark at 544p, PDMD scores higher than the reimplemented DMD, DMD2, rCM and AnyFlow baselines and the H3 Turbo LoRA on several video and audio metrics. It was also tested on Wan2.1-T2V-1.3B with VBench-style prompts. Baselines marked as reimplementations were trained by the authors under a shared protocol, so the comparisons are self-reported.
A YouTube creator mentions that a full modified model of about 66 GB and a 1.4 GB LoRA were released. The provided pages do not give download links, a license, or a paper URL, so those details are unverified here. The work is aimed at researchers and practitioners who want to speed up diffusion models.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Another four-step distillation method that projects away the critic's errors from the training signal. It beats other distillation methods, with a full 66 GB modified model and a 1.4 GB LoRA released.
Closest neighbours by meaning, not just by category.
Research · ByteDance
ByteDance and Texas A&M research method distilling image and video diffusion models into few-step generators adversarially.
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
8-step distillation LoRA for MOSS-Image-SFT that runs without CFG, for ComfyUI.
Research
Diffusion framework that learns text-to-3D human motion generation from 2D poses extracted from monocular videos.
Model · ByteDance
ByteDance's open-weights one-step diffusion transformer for restoring and upscaling video and images.
