Addressing degeneracies in latent interpolation for diffusion models

There is an increasing interest in using image-generating diffusion models for deep data augmentation and image morphing. In this context, it is useful to interpolate between latents produced by inverting a set of input images, in order to generate new images representing some mixture of the inputs. We observe that such interpolation can easily lead to degenerate results when the number of inputs is large. We analyze the cause of this effect theoretically and experimentally, and suggest a suitable remedy. The suggested approach is a relatively simple normalization scheme that is easy to use whenever interpolation between latents is needed. We measure image quality using FID and CLIP embedding distance and show experimentally that baseline interpolation methods lead to a drop in quality metrics long before the degeneration issue is clearly visible. In contrast, our method significantly reduces the degeneration effect and leads to improved quality metrics also in non-degenerate situations.

Tiny models from tiny data: Textual and null-text inversion for few-shot distillation

Few-shot image classification involves classifying images using very few training examples. Recent vision foundation models show excellent few-shot transfer abilities, but are large and slow at inference. Using knowledge distillation, the capabilities of high-performing but slow models can be transferred to tiny, efficient models. However, common distillation methods require a large set of unlabeled data, which is not available in the few-shot setting. To overcome this lack of data, there has been a recent interest in using synthetic data. We expand on this work by presenting a novel diffusion model inversion technique (TINT) combining the diversity of textual inversion with the specificity of null-text inversion. Using this method in a few-shot distillation pipeline leads to state-of-the-art accuracy among small student models on popular benchmarks, while being significantly faster than prior work. This allows us to push even tiny models to high accuracy using only a tiny application-specific dataset, albeit relying on extra data for pre-training. Popular few-shot benchmarks involve evaluation over a large number of episodes, which is computationally cumbersome for methods involving synthetic data generation. Therefore, we also present a theoretical analysis on how the variance of the accuracy estimator depends on the number of episodes and query examples, and use these results to lower the computational effort required for method evaluation. In addition, to further motivate the use of generative models in few-shot distillation, we demonstrate that our method performs better compared to training on real data mined from the dataset used to train the diffusion model. Source code will be made available at https://github.com/pixwse/tiny2.