Intrinsic Distribution Matching improves few-step image distillation
Original title: Intrinsic Distribution Matching
At two network evaluations, IDM scored 94.25% on Chinese and 90.20% on English LongText-Bench.
Intrinsic Distribution Matching keeps the full distribution-matching distillation correction and adds explicit constraints on feature means and per-coordinate standard deviations. Features come from the generator's exponential-moving-average copy, so no external encoder is required. Full-parameter and LoRA experiments use Z-Image-Turbo 6B and Qwen-Image 20B. Ablations attribute gains to the added moment loss.
- Those LongText-Bench scores were the highest among compared methods at that budget.
- Generator EMA features competed with an external DINOv2 encoder on five metrics.
- Experiments cover full-parameter tuning and LoRA on both models.
- Code is available at github.com/inclusionAI/IDM.
- Organizations
- inclusionAI
- AI models
- Z-Image-TurboQwen-Image
From the original article
1 min read at huggingface.co
One-step and few-step text-to-image generation must accurately render the details specified in the prompt despite limited opportunities for refinement. Distribution matching distillation (DMD) uses score differences to bring the generated distribution closer to a reference distribution. Following prior work, we estimate this difference with internal velocity branches, avoiding a separate fake-score network.
By Deyuan Liu, Yansen Han, Kaiwen Zheng · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.