NTH

VA-TDM: A Unified View for Improving Few-Step Diffusion Models by Velocity Adversarial Distillation

AuthorsYihong Luo, Tianyu Pang, Tianyang Hu

AffiliationsHong Kong University of Science and Technology · The Chinese University of Hong Kong, Shenzhen Summarized Prompt: A woman wearing headphones journeys through surreal worlds, ending at a mountain sunrise with the text "hear beyond reality"

November 11, 2026 2 min read
Watch on YouTube
The one-line take

VA-TDM uses adversarial feedback on diffusion velocities to produce high-quality images and videos with just a few generation steps.

Key results

4
Inference steps (NFE)

VA-TDM generation budget for the reported few-step results.

0.67
SD3.5-Medium GenEval

VA-TDM score, compared with TDM's 0.61.

0.69
FLUX.2-Klein-Base OCR similarity

VA-TDM score, compared with 0.64 for the released four-step model.

0.56
FLUX.2-Klein-Base LPIPS diversity

VA-TDM score, compared with 0.47 for the released four-step model.

33B
MiniMax-H3 model scale

Model scale used to demonstrate video-generation distillation.

61.33%
Preference over Fast-H3

Human preference for VA-TDM at the same four-step budget.

What the paper found

VA-TDM addresses a central problem in few-step diffusion: student-score estimates can fail to represent the distributions produced during intermediate denoising. Its underlying Velocity Adversarial Distillation framework aligns teacher and student velocity fields using a learned vector-valued critic. It combines A-form rectified velocities, which provide conditional-moment supervision, with B-form Gaussian velocities, which encode distributional information; together, these connect moment matching, score distillation, and Fisher-based matching. VA-TDM applies this dual-form supervision along partial denoising transitions in the student’s own trajectory, using transition-aware online field estimation and a shared backbone, without adding inference cost. At 4 NFE, it raises SD3.5-Medium’s GenEval score from TDM’s 0.61 to 0.67. On FLUX.2-Klein-Base 4B, it improves OCR similarity from 0.64 to 0.69 and LPIPS diversity from 0.47 to 0.56. The method also scales to 33B MiniMax-H3: in human comparisons, its four-step video generation is preferred over Fast-H3 61.33% of the time, and it receives 45.53% preference against the 100-NFE MiniMax-H3 teacher, compared with 24.67% for Fast-H3.

Original abstract

Few-step distillation offers efficient generation with diffusion and flow models, but preserving quality with only a few sampling steps remains challenging. Distribution-matching methods via score distillation provide a strong foundation, yet they rely on accurately tracking the student's changing distribution. At intermediate generation steps in few-step sampling, student-score estimation can be difficult, and commonly used estimators may describe a different distribution from the one used to update the student. Flow matching allows a more flexible approach to supervision through velocity fields constructed from the student's sampling process. Building on this perspective, we introduce Velocity Adversarial Distillation (VAD), which uses a learned critic to compare student and teacher fields and turn their disagreement into feedback for learning. VAD unifies score distillation, Fisher-based matching, and moment matching, also explaining existing heuristic MMD updates and delivering new training objectives. We further develop VA-TDM, which follows trajectory distribution matching: providing supervision for partial denoising along the student trajectory. Beyond this, VA-TDM combines transition-aware online velocity learning and dual-form VAD supervision across intermediate student transitions. Extensive results show that VA-TDM is powerful for learning few-step generators, achieving state-of-the-art four-step performance in both text-to-image and text-to-video generation across various backbones and scaling effectively to 33B MiniMax-H3. In particular, our VA-TDM-H3 notably outperforms the few-step Fast-H3 and is closer to the original many-step H3 according to human preference.

Read the original paper

More in Diffusion Models

Browse all 82 papers →