Research map / Image generation / Text-to-image diffusion
Beyond Pixel Mimicry: Disentangled Self-Similarity Rewards for Diverse Subject-Driven Generation
Qian Wang, Zhenyu Li, Abdelrahman Eldesokey, Peter Wonka
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DivRL (nSSM + gated VSM) post-training | Flux-Kontext · DreamBench++ | MTG-nSSM: 0.659 -> 0.689 | Improved |
| add DivRL post-training | Flux-Kontext · DreamBench++ | DINO-nSSM: 0.433 -> 0.453 | Improved |
| add DivRL post-training | Flux-Kontext · DreamBench++ | VSM-0.7: 0.605 -> 0.614 | No clear effect |
| add DivRL post-training | Flux-Kontext · DreamBench++ | HPS: 0.294 -> 0.287 | Hurt |
| add nSSM diversity reward to VSM reward (ablation) | DivRL (VSM only) · DreamBench++ | MTG-nSSM: 0.608 -> 0.689 | Improved |
| add nSSM diversity reward to VSM reward (ablation) | DivRL (VSM only) · DreamBench++ | VSM-0.7: 0.688 -> 0.614 | Hurt |
| add VSM identity gate to nSSM-only reward (ablation) | DivRL (nSSM only) · DreamBench++ | VSM-0.7: 0.562 -> 0.614 | Improved |
| add VSM identity gate to nSSM-only reward (ablation) | DivRL (nSSM only) · DreamBench++ | MTG-nSSM: 0.704 -> 0.689 | Hurt |
| add VSM identity gate to nSSM-only reward (ablation) | DivRL (nSSM only) · DreamBench++ | CLIP-T: 0.287 -> 0.279 | Hurt |
| use gated two-stage reward instead of linear weighting of VSM and nSSM (ablation) | DivRL · DreamBench++ | consistency ratio and diversity-over-consistency | Improved |
| add 2x2 average pooling (48x48 -> 24x24) to features for nSSM (ablation) | DivRL · DreamBench++ | visual artifacts | Improved |
Closest papers by meaning
- GWScore: A structural diversity metric for consistent text-to-image generation (NeurIPS 2026 · similarity 0.89)
- WithAnyone: Toward Controllable and ID Consistent Image Generation (ICLR 2026 · similarity 0.89)
- ARGUS: Stacked Multi-View Identity Mosaic Injection for Subject-Preserving Video Generation (NeurIPS 2026 · similarity 0.87)
- FaithfulFaces: Pose-Faithful Facial Identity Preservation for Text-to-Video Generation (NeurIPS 2026 · similarity 0.86)
- MaSC: A Masked Similarity Metric for Evaluating Concept-Driven Generation (NeurIPS 2026 · similarity 0.85)
- Illuminating Visual Identity in Universal Multimodal Embeddings (CVPR 2026 · 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.
