Hugging Face Introduces Correction Training for Diffusion Models: Enhancing Sample Quality and Diversity

Summary:Hugging Face introduces a correction training framework for diffusion models, which introduces a correction step during inference to bridge the training-inference gap and significantly improve sample quality. The method iteratively refines the denoiser's predictions using a corrector trained against the same ground-truth targets. Evaluations across DiT, SiT, and JiT benchmarks demonstrate substantial improvements in sample quality, particularly for unconditional and unguided generation tasks, ou

---

Technical Mechanism Analysis

Diffusion models rely on iterative sampling during generation, but this process can lead to error accumulation due to exposure bias. During training, the denoiser learns from real images interpolated with noise, but during inference, its inputs are constructed from its own imperfect predictions from earlier steps, which it has never encountered during training. This discrepancy can lead to degraded generation quality or the persistence of artifacts.

Hugging Face's correction training method addresses this issue by introducing a corrector that takes the base model's prediction as input and is trained against the same ground-truth targets, turning the correction process into an explicit learning task. During inference, the corrector iteratively refines the denoiser's predictions at each denoising step, preventing early artifacts from persisting into the final sample.

Engineering Trade-offs and Performance

Although the correction process adds to the inference cost, experimental results demonstrate the method's effectiveness across multiple benchmarks:

• Improved Sample Quality: The correction training method significantly improves FID scores for unguided and unconditional generation tasks on models like DiT, SiT, and JiT, showcasing its advantages in generation quality.
• Diversity Preservation: The method maintains the diversity of generated samples while improving quality, avoiding the mode collapse often associated with over-optimization.
• CFG Integration: Combining the correction step with classifier-free guidance (CFG) further enhances diversity and reduces the oversaturation typical of CFG.

Additionally, the research team proposes a method to distill the refined predictions back into the base model to mitigate the additional inference cost. This approach retains the quality improvements from the correction process while avoiding extra inference-time overhead.

Developer Implementation and Deployment Recommendations

For developers, the correction training method offers a straightforward yet effective framework to enhance the generation quality of diffusion models. Here are some deployment recommendations:

1. Model Integration: Integrate the corrector as a component of the diffusion model, dynamically invoking the correction step during inference.
2. Resource Management: Optimize the frequency of correction step invocation based on the application scenario to balance generation quality and inference cost.
3. Distillation Optimization: Use distillation techniques to distill the corrected predictions back into the base model, reducing inference overhead.

Conclusion

Hugging Face's correction training method provides a novel approach to optimizing diffusion models by explicitly bridging the training-inference gap, significantly improving generation quality and diversity. This method holds important implications for both academic research and practical model deployment, offering practical improvements for real-world applications.

---

Source: Hugging Face Daily Papers (2026-10-14)

— END —

Tags: #Hugging Face Follow #Diffusion Models Follow #Correction Training Follow #Sample Quality Follow #Inference Optimization Follow