New FD-loss technique improves autoregressive image generation quality

Researchers have developed a new post-training technique called Fréchet Distributional Post-Training (FD-loss) for autoregressive image generators. This method addresses the mismatch between token-level cross-entropy pre-training and the distributional quality evaluation of generated images. By using representation-space Fréchet distance as the sole objective, FD-loss adapts pretrained discrete generators. This approach has shown significant improvements, reducing FID and FD_r6 scores by an average of 41.4% and 52.0% respectively across various configurations and datasets. AI

IMPACT Introduces a novel post-training method that significantly enhances the quality of generated images by addressing objective mismatches in autoregressive models.

RANK_REASON Academic paper detailing a new method for image generation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• FD-loss
• Fréchet Distributional Post-Training
• ImageNet

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →