Research map / Image generation / Text-to-image diffusion
Fine-Tuning Visual Autogressive Models for Subject-Driven Generation
Jiwoo Chung, Sangeek Hyun, Hyunjun Kim, Eunseo Koh, MinKyu Lee, Jae-Pil Heo
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 |
|---|---|---|---|
| tune only CA and FFN layers instead of all layers (ablation) | ARBooth (Infinity-2B) · 5 living subjects | PRES / DIV: PRES 0.776 -> 0.709; DIV 0.254 -> 0.272 | Improved |
| add scale-wise weighted tuning (down-weight scales above 8 to 0.5) (ablation) | ARBooth · 5 living subjects | I_dino: 0.774 -> 0.786 I_dino; PRES 0.792 -> 0.709; T_clip 0.271 -> 0.267 | Improved |
| add prior distillation (ablation) | ARBooth · 5 living subjects | PRES / DIV: PRES 0.839 -> 0.709; DIV 0.209 -> 0.272; I_dino 0.823 -> 0.786 | Improved |
| replace prior distillation with prior preservation loss (class images) (ablation) | ARBooth · 5 living subjects | I_dino / DIV: I_dino 0.786 -> 0.653; DIV 0.272 -> 0.378 | No clear effect |
| use LoRA on all layers without selective layer tuning or prior distillation (ablation) | ARBooth · 5 living subjects | subject fidelity: LoRA(4) I_dino 0.653 vs 0.786 | Hurt |
| use ARBooth on Infinity-2B instead of diffusion-based personalization baselines | subject-driven generation · ViCo dataset | I_dino: 0.682 (DM) -> 0.705 | Improved |
| use ARBooth instead of DreamBooth on recent diffusion models (SD3.5, FLUX) | subject-driven generation · 5 living subjects | I_dino: 0.730 (SD3.5-md Full) -> 0.786; time 15s -> 0.5s | Improved |
| use ARBooth instead of SD3.5-medium LoRA | subject-driven generation · 5 living subjects | T_clip: 0.275 -> 0.267 | Hurt |
Closest papers by meaning
- OneActor: Consistent Subject Generation via Cluster-Conditioned Guidance (NeurIPS 2024 · similarity 0.89)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.89)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.89)
- TTS-VAR: A Test-Time Scaling Framework for Visual Auto-Regressive Generation (NeurIPS 2025 · similarity 0.89)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 2025 · similarity 0.89)
- Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective (NeurIPS 2024 · similarity 0.88)
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.
