TLDR: DMIS is a novel framework that enables conditional diffusion models to learn effectively from imprecise supervision, such as noisy, ambiguous, or incomplete labels. It decomposes the learning objective into generative and classification components, using a weighted denoising score matching for label-conditioned learning and an efficient diffusion classifier for inferring class probabilities. Experiments show DMIS consistently produces high-quality, class-discriminative samples across image generation, weakly supervised learning, and noisy dataset condensation, pioneering robust learning in real-world imperfect data scenarios.
Conditional diffusion models have recently achieved remarkable success in various generative tasks, demonstrating unprecedented capabilities in creating realistic data. These models are particularly powerful when conditioned on additional information, such as text descriptions or class labels, enabling tasks like text-to-image synthesis and image inpainting.
However, a significant challenge arises in real-world scenarios: the conditioning information required by these models is often imprecise. This imprecision can stem from noisy, ambiguous, or incomplete labels, frequently encountered when data is sourced from the internet or through crowdsourcing. Such imperfect supervision can lead to a mismatch between the input conditions and the true underlying labels, ultimately degrading the quality of generated content.
To tackle this pervasive issue, researchers have introduced DMIS, a unified framework designed for training robust Diffusion Models from Imprecise Supervision. This framework represents the first systematic study within the field of diffusion models to address this problem comprehensively. DMIS is rooted in the principle of likelihood maximization, breaking down the overall objective into two core components: a generative component and a classification component.
The generative component of DMIS is responsible for modeling distributions based on imprecise labels. It achieves this by expressing the imprecise-label conditional score as a combination of clean-label conditional scores, weighted by their posterior probabilities. This insight led to the development of a weighted denoising score matching objective, which allows the model to learn label-conditioned generation without needing perfectly clean annotations.
The classification component, on the other hand, leverages a diffusion classifier to infer class-posterior probabilities. This is crucial for understanding the true labels hidden within imprecise supervision. To enhance efficiency, DMIS incorporates an optimized timestep sampling strategy for this diffusion classifier. This strategy significantly reduces computational costs without compromising performance, by identifying a compact subset of timesteps that are most informative for classification.
Extensive experiments were conducted across diverse forms of imprecise supervision, covering critical tasks such as image generation, weakly supervised learning, and a novel area called noisy dataset condensation. The results consistently demonstrated that DMIS produces high-quality and class-discriminative samples, outperforming existing methods. For instance, in image generation, DMIS not only improved sample quality but also generated distributions that more closely matched the true data distribution, even under challenging conditions like partial-label supervision.
In weakly supervised learning scenarios, DMIS, evaluated through its diffusion classifier, achieved superior performance compared to prior discriminative approaches, highlighting the stronger generalization capabilities of diffusion models. Furthermore, DMIS pioneered the study of noisy dataset condensation, a practical yet previously unexplored setting. It consistently surpassed other methods in condensing clean datasets from noisy-label data, implicitly denoising labels during generation rather than memorizing noisy ones.
Also Read:
- MaskGRPO: A Unified Reinforcement Learning Approach for Multimodal Discrete Diffusion Models
- Unlocking Compositional Generalization in AI Image and Video Creation
This work establishes a solid baseline for future research in robust and scalable diffusion modeling under weak supervision. By providing a unified and efficient framework, DMIS addresses a critical real-world challenge, enabling the development of more reliable generative AI systems even when perfect data is unavailable. For more details, you can refer to the full research paper here.
