Research map / Video / Video generation
MorphSIG: Subject-Driven Image Generation via Decoupled Anchoring and Feature Transport
Hailong Yan, Yongrui Zhang, Xiangtao Zhang, Le Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MorphSIG, a training-free subject-driven image generation framework that reformulates SIG as pseudo-video feature transport with progressive decoupled anchoring and pseudo-flow guidance.
Problem
Subject-driven image generation must preserve a reference subject's identity while composing it into new scenes; the abstract indicates existing methods suffer from structural locking during early denoising.
Approach
Reformulates subject-driven generation as pseudo-video feature transport; Progressive Decoupled Anchoring uses frequency-domain decoupling and structural scrambling in early denoising to break structural locking and transition from free composition to identity alignment, paired with Pseudo-flow Guidance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Progressive Decoupled Anchoring (ablation) | MorphSIG · ViStoryBench ablation | CIDS: 0.614 -> 0.629 | Improved |
| add structural permutation to PDA (ablation) | PDA · ViStoryBench ablation | CIDS: 0.556 -> 0.629 | Improved |
| add spectral decoupling to PDA (ablation) | PDA · ViStoryBench ablation | INC: 13.57 -> 13.88 | Improved |
| use high-pass only instead of full PDA (ablation) | PDA · ViStoryBench ablation | CIDS: 0.629 -> 0.485 | Hurt |
| add pseudo-flow guided feature transport (WFT) (ablation) | MorphSIG · ViStoryBench ablation | CIDS: 0.541 -> 0.629 | Improved |
| add mutual nearest neighbor matching (ablation) | WFT · ViStoryBench ablation | CIDS: 0.578 -> 0.629 | Improved |
| add low-frequency SLERP (ablation) | WFT · ViStoryBench ablation | CIDS: 0.584 -> 0.629 | Improved |
| add high-frequency winner-take-all (ablation) | WFT · ViStoryBench ablation | AES: 5.14 -> 5.39 | Improved |
| replace frequency-decoupled transport with direct ADD (ablation) | WFT · ViStoryBench ablation | CIDS: 0.629 -> 0.561 | Hurt |
| remove WFT or SP: prompt alignment rises (ablation) | MorphSIG · ViStoryBench ablation | PA: w/o SP 2.72, w/o WFT 2.64 vs full 2.58 | Mixed |
| Local Consensus neighborhood K=4 (ablation) | WFT · ViStoryBench | CIDS and INC | Improved |
| extend PDA beyond 20% of denoising steps (ablation) | PDA schedule · ViStoryBench | detail alignment | Hurt |
Closest papers by meaning
- CoDi: Subject-Consistent and Pose-Diverse Text-to-Image Generation (ICLR 2026 · similarity 0.88)
- OpenSubject: Leveraging Video-Derived Identity and Diversity Priors for Subject-driven Image Generation and Manipulation (ECCV 2026 · similarity 0.87)
- Less-to-More Generalization: Unlocking More Controllability by In-Context Generation (ICCV 2025 · similarity 0.86)
- MorphGS: Morphology-Adaptive Articulated Motion Transfer from Videos (ECCV 2026 · similarity 0.86)
- Reconstruct, Inpaint, Test-Time Finetune: Dynamic Novel-view Synthesis from Monocular Videos (NeurIPS 2025 · similarity 0.85)
- Emergent Temporal Correspondences from Video Diffusion Transformers (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
