A new AI training method makes text-to-motion models dramatically smaller and faster without touching real motion-capture data.
Researchers introduce TACD (Terminal-Amplification-Controlled Distillation), which builds on segmented on-policy flow distillation to train compact "student" motion generators from pretrained "teacher" models. The method fixes a specific flaw in prior distillation techniques: when matching velocity predictions on a fixed supervision grid, errors near the end of the denoising process get overweighted, which hurts quality when a model has to generate motion in very few steps. TACD ties the teacher's final query to the student's own step size, capping how much weight those endpoint errors get without changing how the model runs at inference time. On the HumanML3D and KIT-ML benchmarks, eight-step students trained with TACD cut FID error (a standard measure of generation quality) by 58% compared to distillation without the fix, and four-step students matched or beat the quality of their 50-step diffusion teachers.
Text-to-motion models - used to animate game and film characters, VR avatars, and robotics simulations from a text prompt - have followed the same path as image diffusion models: accurate but slow, many-step generation that's expensive to run. TACD's eight-step students hit 7.7-11.9x faster inference and used 3.8-6.7x less GPU memory than their teachers, the kind of efficiency gain that decides whether motion generation runs on a render farm or inside a game engine in real time.
It's the same compression bet that made fast-diffusion tricks useful for image generation - and until TACD shows up inside an actual animation tool, the speedup numbers are a lab result, not a shipped product.