On Distillation of Guided Diffusion Models

Implements a two-stage pipeline that first trains a single student model to match the combined output of separate class-conditional and unconditional teacher models (Stage 1: Output Matching), then progressively distills the matched model to reduce required denoising steps from 50-100+ to 1-4 steps (Stage 2: Progressive Distillation). The approach preserves classifier-free guidance by matching the guidance-weighted output formula: p_θ(x|y) + w(p_θ(x|y) - p_θ(x)), enabling knowledge transfer while maintaining generation quality as measured by FID/IS metrics.

Enables fast text-to-image generation using distilled diffusion models that require only 1-4 denoising steps instead of 50-100+ steps. The capability leverages the two-stage distillation pipeline to compress guidance information into a single efficient model, maintaining semantic alignment between text prompts and generated images while reducing inference latency. Tested on LAION-scale datasets and latent-space architectures (e.g., Stable Diffusion).

Enables efficient image editing by applying text-guided diffusion with only 2-4 denoising steps instead of 50+ steps. The capability leverages distilled models to perform semantic image modifications (e.g., style transfer, object replacement, attribute editing) while preserving unedited regions. Works by conditioning the diffusion process on both the original image and text instructions, using the compressed guidance mechanism from the two-stage distillation pipeline.

Performs image inpainting (filling masked regions) using distilled diffusion models with 1-4 denoising steps. The capability leverages the two-stage distillation pipeline to compress guidance information while maintaining semantic coherence in inpainted regions. Works by conditioning the diffusion process on the original image, inpainting mask, and optional text guidance, enabling fast content-aware region filling without retraining.

Applies the two-stage distillation pipeline to pixel-space diffusion models (operating directly on image pixels rather than latent representations). The capability reduces sampling steps from 50+ to 4 steps while maintaining FID/IS metrics on datasets like ImageNet 64x64 and CIFAR-10. Pixel-space distillation is computationally more expensive than latent-space but provides direct pixel-level control and interpretability.

Applies the two-stage distillation pipeline to latent-space diffusion models (operating on VAE-encoded representations). The capability reduces sampling steps to 1-4 steps while maintaining FID/IS metrics on high-resolution datasets (ImageNet 256x256, LAION). Latent-space distillation is computationally efficient and achieves 10-256× speedup by compressing the guidance mechanism within the VAE latent space, enabling fast inference on resource-constrained hardware.

Implements Stage 2 of the distillation pipeline: iteratively reducing required denoising steps from the output-matched model (typically 50+ steps) down to 1-4 steps through sequential distillation rounds. Each round trains a new student model to match the previous model's output with fewer steps, enabling gradual compression without catastrophic quality collapse. The approach preserves FID/IS metrics across reduction stages by carefully balancing step reduction rate and training data.

Implements Stage 1 of the distillation pipeline: training a single student model to replicate the combined output of separate class-conditional and unconditional teacher models. The student learns to match the guidance-weighted output formula: p_θ(x|y) + w(p_θ(x|y) - p_θ(x)), where w is the guidance scale. This stage consolidates two teacher models into one efficient student while preserving the guidance mechanism, enabling subsequent progressive distillation without guidance degradation.

Monitors and preserves Fréchet Inception Distance (FID) and Inception Score (IS) metrics throughout the two-stage distillation pipeline. The approach ensures that output-matched models and progressively distilled models maintain comparable FID/IS scores to original models, providing quantitative evidence that generation quality is preserved despite step reduction. Metrics are computed on standard benchmarks (ImageNet, CIFAR-10, LAION) to enable comparison across architectures and datasets.

Applies the two-stage distillation pipeline across diverse datasets (ImageNet 64x64, CIFAR-10, ImageNet 256x256, LAION) with dataset-specific hyperparameter tuning. The approach demonstrates that distillation effectiveness varies by dataset characteristics (resolution, diversity, caption quality), enabling practitioners to optimize distillation for their specific data distribution. Latent-space distillation on LAION achieves 1-4 steps while maintaining quality on large-scale text-image data.