Research map / Image generation / Text-to-image diffusion
Don’t Settle at the Mode! Mitigating Diversity Collapse in Pretrained Flow Models via Feature Self-Guidance
Pradhaan S Bhat, Rishubh Parihar, Abhijnya Bhat, Venkatesh Babu Radhakrishnan
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 feature self-guidance with manifold regularization | FLUX.1-dev · GenEval T2I | DINO diversity: 0.57 -> 0.68 | Improved |
| add feature self-guidance | FLUX.1-dev · GenEval T2I | CLIPScore: 32.43 -> 32.05 | Hurt |
| add feature self-guidance, depth-conditioned | FLUX.1-Depth-dev · COCO depth-to-image | DINO diversity: 0.43 -> 0.52 | Improved |
| add feature self-guidance, reference-conditioned | FLUX.1-Kontext-dev · DreamBooth | DINO diversity: 0.45 -> 0.57 | Improved |
| add feature self-guidance | Qwen-Image · GenEval | DINO diversity: 0.65 -> 0.67 | Improved |
| add feature self-guidance | Sana1.5 · GenEval | DINO diversity: 0.45 -> 0.50 | Improved |
| move guidance from block B2 to B1 (ablation) | feature self-guidance · GenEval T2I | CLIPScore: 32.05 -> 31.29 | Mixed |
| move guidance from block B2 to B3 (ablation) | feature self-guidance · GenEval T2I | DINO diversity: 0.68 -> 0.60 | Hurt |
| stop guidance at t=0.9 instead of 0.8 (ablation) | feature self-guidance · GenEval T2I | DINO diversity: 0.68 -> 0.66 | Hurt |
| extend guidance to t=0.7 (ablation) | feature self-guidance · GenEval T2I | DINO diversity: 0.68 -> 0.68 | No clear effect |
| add manifold regularization with higher beta (ablation) | feature self-guidance · text-to-image | prompt adherence | Mixed |
| use Particle Guidance instead of IID | FLUX.1-dev · GenEval T2I | DINO diversity: 0.57 -> 0.58 | No clear effect |
Closest papers by meaning
- DiverseFlow: Sample-Efficient Diverse Mode Coverage in Flows (CVPR 2025 · similarity 0.91)
- Improved Training Technique for Shortcut Models (NeurIPS 2025 · similarity 0.89)
- Efficient Dataset Distillation for Pre-Trained Self-Supervised Models via Statistical Flow Matching (NeurIPS 2026 · similarity 0.87)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.87)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.87)
- Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation (CVPR 2025 · similarity 0.87)
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.
