Generative models for interactive content creation and understanding in smart imaging

Generative diffusion models have achieved substantial progress in image content creation, yet practical interactive generation remains challenging when heterogeneous conditions must be jointly interpreted, fine-grained attributes must be controlled independently, and users need to iteratively refine generated results. This paper proposes a Multimodal Interactive Generation Framework (MIGF) that integrates three task-oriented components: Adaptive Condition Fusion (ACF), Semantic Decoupling (SDM), and an Interactive Feedback Mechanism (IFM). ACF addresses sample-dependent reliability differences among text, reference-image, and structured conditions by predicting modality-specific fusion weights for each input. SDM uses contrastive supervision to organize latent representations into content- and style-related subspaces, facilitating more independent attribute control. IFM integrates local modification, attribute adjustment, and example guidance into a unified progressive refinement process. We construct a multimodal dataset containing 50,000 high-resolution images covering natural scenes, artistic works, design patterns, and architecture. Under matched evaluation settings, MIGF reduces FID by 23.7%, improves CLIP Score by 18.6%, and reduces LPIPS by 12.7% relative to the strongest self-run baseline, ControlNet. Ablation experiments show consistent contributions from the three components. The combination of text, image, and sketch conditions achieves a Control Precision of 0.876, corresponding to an 11.9% improvement over the strongest single-modality configuration. In addition, the fast inference mode reduces generation time to 1.9 s per image. These results indicate that the proposed framework provides a practical approach to multimodal and interactively controllable image generation, while its remaining limitations in robustness, computational cost, and cross-domain generalization motivate further investigation.