Oct 2026· International Conference on Computer Vision, Robotics and Automation Engineering· Vol 14353, pp. 143530C - 143530C-8· 0 citations· 23 references
Engineering
TL;DR
Experimental results verify that classifier-free guidance effectively outperforms traditional conditional diffusion methods in overall generation performance, and proper hyperparameter combination can further narrow the distribution gap between synthetic images and real scene data.
Abstract
Diffusion models have achieved remarkable progress in visual generation, yet their practical deployment for multi-class procedural scene synthesis still lacks standardized training schemes and comprehensive performance verification. This study adopts DDPM as the base architecture and introduces classifier-free guidance to facilitate controllable multi-category image generation toward Procgen game scenes. We formulate unified data preprocessing rules and consistent training settings, and quantitatively compare the generation capability of unconditional diffusion, vanilla class-conditional diffusion, and classifier-free guided diffusion models via FID metrics. We further explore the influence of core hyperparameters, including label drop probability and guidance scale on generated sample quality, fidelity, and diversity. Experimental results verify that classifier-free guidance effectively outperforms traditional conditional diffusion methods in overall generation performance, and proper hyperparameter combination can further narrow the distribution gap between synthetic images and real scene data. In addition, we observe obvious performance differences across categories with distinct visual complexity. This work provides solid empirical evidence and practical tuning experience for applying conditional diffusion models to procedural multi-class image generation, and opens feasible directions for subsequent adaptive conditional generation research.
Swift-Image achieves leading aggregate performance among evaluated open-source models with only 6B parameters and 243K GPU training hours; the compressed 3B model incurs nearly no loss, while few-step distillation further improves aggregate editing performance with substantially fewer sampling steps.
The practical success of conditional image generation hinges on fine-grained differences in condition alignment and visual fidelity. Classifier-free guidance (CFG) is central to this success, but its lack of an explicit criterion makes it difficult to assess whether the guided trajectory is progressing as intended. To...
In recent years, the progress of diffusion models in text-to-image generation has been obvious to all. Text-based descriptions alone, however, are often insufficient for precisely controlling screen content. In order to solve this problem, researchers began to try to introduce additional conditions or visual references...
Recent advances in deep learning for shadow removal have significantly enhanced image quality and realism. However, most approaches rely on real-world paired datasets, which are costly to collect and often limited in scene diversity, leading to limited generalization. To address these limitations, we propose One-step D...
Junseong Shin, Kijun Kim, Minseong Kim et al.· 0 citations
GoM, short for Generalization over Memorization, the framework combines scene-disjoint validation, exposure-matched selection, and targeted diffusion adaptation in a single-Image Guided Multi-Angle Image Synthesis challenge.
This paper systematically sorts out two representative controllable generation methods: the explicit structural constraint method based on additional conditions and the implicit semantic guidance method based on visual examples, which realizes precise control of image spatial layout and geometric structure.
Bing-Yi Zhao· ITM Web of Conferences· 0 citations
Related blog posts
Microsoft Research Blog· microsoft.comNov 16, 2026
AI coding agents are increasingly help developers fix bugs and build features largely on their own, but every interaction with their models consumes tokens. In each session, an agent sends instructions, tool descriptions, and session history to its model. Most efforts to reduce this cost focus on individual tasks. Yet developers return to the same […] The post TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient appeared first on Microsoft Research.
MIT scientists developed a training method to generate models that simulate the effects of human activities on climate much more quickly than other approaches.