Few-step generation leaves little room to correct missing prompt details, so a useful distillation objective needs to do more than steer samples toward a target distribution. The key idea here is to add explicit constraints on feature statistics without replacing the learned score correction that already drives distribution matching.
Key Findings
- The method preserves the full DMD correction and adds a Gaussian moment constraint on feature means and per-coordinate standard deviations. This supplements the learned signal rather than narrowing it to moment matching alone.
- Features come from the generator’s exponential-moving-average copy, eliminating the need for a separate feature encoder in the proposed setup.
- At two network evaluations, it reports LongText-Bench scores of 94.25% for Chinese and 90.20% for English, the highest among the compared methods at that evaluation budget. These results are specific to the reported experiments, which tune Z-Image-Turbo 6B and Qwen-Image 20B using full-parameter or LoRA adaptation.
Who Should Care—and When Not To
A good fit for researchers working on text-to-image distillation who want to improve prompt-detail retention while keeping score-based supervision intact. The evidence is tied to the paper’s models, tuning setups, and LongText-Bench evaluation; it does not establish gains across all generators or image-generation benchmarks.
