MorphSIG improves subject consistency in image generation with new feature transport
MorphSIG: Subject-Driven Image Generation via Decoupled Anchoring and Feature Transport
Computer Vision and Pattern Recognition
Summary
Creating images of specific subjects in different poses usually struggles to keep the subject looking consistent. The authors developed MorphSIG, a method that treats this as moving features like a video but without needing extra training. They use clever strategies to mix and match subject identity and pose progressively, producing more varied and faithful images. They also created a new test set with diverse styles and found their method works better than others without extra learning.
What this means in practice
- •For graphic designers: Generate diverse images of a specific subject in different poses with consistent identity for creative projects.
- •For animation studios: Create coherent sequences of stylized subject images without requiring video data or extra training.
Authors
Hailong Yan, Yongrui Zhang, Xiangtao Zhang, Le Zhang
Abstract
We propose MorphSIG, a training-free framework that reformulates subject-driven image generation (SIG) as pseudo-video feature transport. MorphSIG contains two core modules: Progressive Decoupled Anchoring, which uses frequency-domain decoupling and structural scrambling during early denoising to break structural locking and enable a gradual transition from free composition to identity alignment; and Pseudo-flow Guided Feature Transport, which constructs a latent-space pseudo-flow field to mimic Image-to-Video temporal coherence and transport subject semantics precisely. To address stylistic homogeneity and the lack of ground truth in existing benchmarks, we further introduce a benchmark with diverse stylized subjects. Extensive experiments show that MorphSIG significantly improves pose diversity and subject fidelity over baselines without additional training. Overall, MorphSIG bridges static generation and dynamic propagation, providing a cost-effective paradigm for high-fidelity subject consistency.