IDM boosts few-step text-to-image generation
In brief
- IDM adds statistical moment constraints to DMD for improved few-step text-to-image generation.
- Scores reach 94.25% and 90.20% on LongText-Bench in Chinese and English with two network evaluations.
- Generator EMA features replace an external encoder without performance loss.
Researchers introduce Intrinsic Distribution Matching (IDM), a method to enhance few-step text-to-image generation. According to the announcement, IDM adds explicit statistical moment constraints to distribution matching distillation (DMD) to better align the generated distribution with the reference.
IDM retains the full DMD correction and leverages the generator’s EMA features to constrain means and per-coordinate standard deviations, without an external encoder. Experiments on Z-Image-Turbo 6B and Qwen-Image 20B with full or LoRA tuning report scores of 94.25% and 90.20% on LongText-Bench in Chinese and English at two network evaluations. Controlled ablations confirm gains from the moment loss, while EMA features match DINOv2 performance across five metrics.
This advance targets fast image generation models, where rendering long prompts accurately remains challenging with limited evaluations. Code is open source, and model weights will be released soon.
What it changes for you
- Media & EntertainmentContent creators and studios can generate more accurate visuals at lower compute cost, improving production quality with lightweight models.
This summary gives you the essentials. The full article, its figures and quotes are on Hugging Face Daily Papers.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error