Research map / Image generation / Text-to-image diffusion
Proxy-Tuning: Tailoring Multimodal Autoregressive Models for Subject-Driven Image Generation
Yi Wu, Shengju Qian, Lingting Zhu, Lei Liu, Wandi Qiao, Ziqiang Li, Lequan Yu, Bin Li
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 |
|---|---|---|---|
| fine-tune AR model on diffusion-supervisor proxy images (Proxy-Tuning) instead of direct LoRA | LlamaGen · DreamBooth (SDXL supervisor) | DINO: 0.5088 -> 0.7436 | Improved |
| Proxy-Tuning instead of direct LoRA | Lumina-mGPT · DreamBooth (SDXL supervisor) | DINO: 0.4504 -> 0.7834 | Improved |
| AR student vs diffusion supervisor on subject fidelity | Lumina-mGPT · DreamBooth (SDXL supervisor) | CLIP-I: 0.8002 -> 0.8074 | Improved |
| Proxy-Tuning on prompt adherence (CLIP-T) | LlamaGen · DreamBooth (SDXL supervisor) | CLIP-T: 0.3225 -> 0.2772 | Hurt |
| Proxy-Tuning on prompt adherence (CLIP-T) | Lumina-mGPT · DreamBooth (SDXL supervisor) | CLIP-T: 0.3101 -> 0.3118 | Mixed |
| apply Proxy-Tuning to diffusion students instead of AR (ablation) | SDXL · DreamBooth | DINO: 0.7272 -> 0.6890 | Hurt |
| apply Proxy-Tuning to diffusion students (ablation) | SD3 · DreamBooth | CLIP-I: 0.7822 -> 0.7596 | Hurt |
| learn multiple subjects jointly in one AR model | LlamaGen (SDXL) · Multi-subject DreamBooth | DINO: 0.7040 -> 0.7176 | Improved |
| Proxy-Tuning vs fine-tuning AR on prompt fidelity (human) | AR model · User study | Prompt fidelity: 2.98 -> 4.52 | Improved |
| Proxy-Tuning for FLUX supervisor on LlamaGen subject fidelity | LlamaGen · DreamBooth (FLUX supervisor) | CLIP-I: 0.8023 -> 0.7974 | No clear effect |
| vary number of proxy images and LoRA rank (ablation) | LlamaGen · DreamBooth (SDXL supervisor) | CLIP-I / DINO | No clear effect |
| geometric data augmentation instead of proxy data (ablation) | AR model · DreamBooth | semantic editability | Hurt |
Closest papers by meaning
- Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.90)
- OneActor: Consistent Subject Generation via Cluster-Conditioned Guidance (NeurIPS 2024 · similarity 0.90)
- Jointly Training Large Autoregressive Multimodal Models (ICLR 2024 · similarity 0.89)
- Image Regeneration: Evaluating Text-to-Image Model via Generating Identical Image with Multimodal Large Language Models (AAAI 2025 · similarity 0.88)
- Diffuse Everything: Multimodal Diffusion Models on Arbitrary State Spaces (ICML 2025 · similarity 0.88)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · 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.
