VA-TDM reports 4-step wins on FLUX.2-Klein and MiniMax-H3
TL;DR
- On FLUX.2-Klein-Base 4B at four steps, VA-TDM reports GenEval 0.80 vs 0.78, OCR similarity 0.69 vs 0.64, and LPIPS diversity 0.56 vs 0.47.
- On 33B MiniMax-H3 at a 4-NFE budget, a user study has VA-TDM-H3 winning 61.33% of pairwise comparisons against Fast-H3 at the same budget.
- The framework unifies score distillation, Fisher-based matching, and moment matching as an adversarial game between student and teacher velocity fields.
On FLUX.2-Klein-Base 4B at four sampling steps, VA-TDM reports a GenEval score of 0.80 against the released four-step model's 0.78, with OCR similarity rising from 0.64 to 0.69 and LPIPS diversity from 0.47 to 0.56.
The paper on Hugging Face, titled 'VA-TDM: A Unified View for Improving Few-Step Diffusion Models by Velocity Adversarial Distillation,' introduces Velocity Adversarial Distillation, which 'uses a learned critic to compare student and teacher fields and turn their disagreement into feedback for learning.' The authors write that the framework 'unifies score distillation, Fisher-based matching, and moment matching, also explaining existing heuristic MMD updates and delivering new training objectives.'
Scaled to the 33B text-to-video MiniMax-H3 at a 4-NFE budget, VA-TDM-H3 'wins 61.33% of comparisons against Fast-H3 at same 4-NFE budget,' and reaches a 45.53% preference rate against the 100-NFE teacher, versus 24.67% for Fast-H3. The abstract claims 'state-of-the-art four-step performance in both text-to-image and text-to-video generation across various backbones.'
It lands amid a steady run of AI video distillation and RL-training papers we have tracked this week. Code is at github.com/Luo-Yihong/VA-TDM. The fetched page lists no author affiliations.
Originally reported by huggingface.co
Read the original article →Original headline: VA-TDM Unifies Score, Fisher and Moment Matching Under Velocity Adversarial Distillation