PDMD open-sources 4-step video distillation for MiniMax-H3

PDMD is a new open-source video distillation method that slashes MiniMax-H3 generation steps from the usual 20 to 30 down to just four.

It makes running video models locally significantly faster while sidestepping the muddy, washed-out artifacts that typically plague aggressive distillation.

What it does

Normally, generating a video clip with MiniMax-H3 requires between 20 and 30 denoising steps (NFE), which demands serious GPU horsepower and testing patience.

Applying PDMD compresses that entire workflow into just four steps. Instead of waiting minutes for a render, frames start churning out in a fraction of the time.

Crucially, visual fidelity holds up surprisingly well. Compared side-by-side with other 4-step distillation techniques, PDMD produces videos with noticeably sharper textures, steadier subject consistency, and far cleaner handling of complex motion. On standard video and audio quality benchmarks, it consistently outperforms existing alternatives.

How it works

PDMD tackles a fundamental flaw in how student diffusion models are trained: flawed refereeing.

In standard distillation setups, a Critic network acts as an evaluator, comparing the student model's output against the full-size teacher model and sending correction signals back to the student. The catch is that the critic itself is imperfect. When the critic makes a bad call, its own hallucinations get baked directly into the student model, turning outputs noisy or blurry over time.

PDMD, short for Projected Distribution Matching Distillation, isolates the portion of the training gradient caused by the critic's own errors and projects it away. By filtering out bad advice before it reaches the student, the model retains the teacher's original motion dynamics and fine details even in ultra-short sampling runs.

How to run it

The research team released code, an arXiv preprint, and model checkpoints on Hugging Face across two formats.

The first option is a full-replacement transformer model (pdmd_4NFE_full) sitting at roughly 66GB. It replaces the core 3D transformer inside MiniMax-H3 for direct 4-step sampling if you have serious VRAM to spare.

The second option is a lightweight LoRA adapter (pdmd_4NFE_lora) weighing in at just under 1.4GB. You can pair this adapter with a quantized MiniMax-H3 base model, leaving the text encoder and VAE untouched. It loads easily inside Hugging Face Diffusers with standard pipeline calls, and settings are already provided for local apps like Draw Things as well as Google Colab notebooks.

Why it matters

For local AI creators and developers, generation latency has always been the real bottleneck. Spending minutes per generation kills prompt tweaking; cutting that cycle down to four steps on a modest 1.4GB adapter makes local video experiments practical on reasonable hardware.

More broadly, PDMD offers an elegant algorithmic takeaway. It shows that extreme-speed distillation does not have to sacrifice image sharpness—the quality drop was largely due to flawed feedback from imperfect critics. If this error-projection logic catches on across other open weights, open-source video generation will get dramatically faster very quickly.