SMD few-step diffusion distillation: no score model or GAN, validated up to 33B, fully released
ArashVahdat · x · 2026-10-22
Researchers released Simplified Matching Distillation (SMD) for few-step diffusion: no auxiliary score model, no adversarial training, no generator modification. It yields photorealistic appearance, high-quality dynamics, memory-efficient training, and mode-covering diverse outputs—validated across images, video, joint audio–video, and 3D up to 33B parameters. It also supports multi-teacher capability merging (like OPD in LLMs), cross-model capability transfer, and bidirectional teacher-to-causal distillation, with paper, code, models, and data fully released.
NVIDIA's Julius Berner notes SMD is very close to their concurrent Parametric Trajectory Distillation (PTD, EPFL/NVIDIA/Stanford, demoed on MiniMax-H3): where PDD uses a piecewise-linear student trajectory, SMD/PTD use basis expansions with learned coefficients discarded at inference, and all three can be seen as special cases of Lagrangian flow-maps with closed-form JVPs.
More from Multimodal
- 16 AI tips to make a full music video with just 4 tools — techhalla · 2026-10-22
- HeyGen's video and voice models land on fal: built-in sound, 39-language voice cloning — HeyGen · 2026-10-22
- Filmmaker uses Grok for movie previs: strong art direction, sync still lacking — TinfoilTricorn · 2026-10-22
- Dev builds lightweight real-time AI conversational avatars, native in Unreal Engine 5 — Sad_Eagle_937 · 2026-10-22
- 3D-plus-AI video workflow prediction from 2024 has become the default, says Bilawal Sidhu — nptacek · 2026-10-22
- Redditor recreates MC Standly's music video with MiniMax H3 — pwillia7 · 2026-10-22