Research map / Vision-language and multimodal / Multimodal LLMs
GenHancer: Imperfect Generative Models are Secretly Strong Vision-Centric Enhancers
Shijie Ma, Yuying Ge, Teng Wang, Yuxin Guo, Yixiao Ge, Ying Shan
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 |
|---|---|---|---|
| enhance CLIP with GenHancer instead of DIVA (lightweight denoiser) | OpenAI ViT-L-14 @224 · MMVP-VLM | average accuracy: 25.9 -> 31.9 | Improved |
| enhance CLIP with GenHancer instead of DIVA | SigLIP ViT-SO-14 @224 · MMVP-VLM | average accuracy: 40.7 -> 42.2 | Improved |
| use enhanced CLIP in LLaVA-1.5 versus original CLIP | Vicuna-7B · MMVP-MLLM | accuracy: 24.7 -> 30.7 | Improved |
| use enhanced CLIP in LLaVA-1.5 versus original CLIP | Vicuna-7B · CV-Bench 3D | accuracy: 58.7 -> 63.2 | Improved |
| use GenHancer CLIP versus DIVA CLIP in LLaVA-1.5 | Vicuna-7B · MMVP-MLLM | accuracy: 31.3 -> 30.7 | Hurt |
| post-tune CLIP with GenHancer on global semantics | OpenAI ViT-L-14 @224 · ImageNet-1K zero-shot | accuracy: 75.5 -> 75.6 | No clear effect |
| use scaled Logit-Normal timestamp sampling (s=1) instead of uniform (ablation) | continuous denoiser · MMVP-VLM OpenAI@224 | average accuracy: 21.5 -> 31.9 | Improved |
| use too large scale s=10 in timestamp sampling instead of s=1 (ablation) | continuous denoiser · MMVP-VLM OpenAI@224 | average accuracy: 31.9 -> 20.7 | Hurt |
| use two-stage training instead of end-to-end training (ablation) | GenHancer · MMVP-VLM | average accuracy | Improved |
| condition only on the class token instead of adding local tokens (ablation) | GenHancer · MMVP-VLM | average accuracy | Improved |
| use LoRA instead of direct training of CLIP in stage 2 (ablation) | GenHancer · MMVP-VLM | average accuracy | Improved |
| use lightweight denoiser instead of pre-trained FLUX.1-dev (ablation) | OpenAI ViT-L-14 @224 · MMVP-VLM | average accuracy: 32.6 -> 31.9 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports GenHancer as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Guiding Diffusion-based Reconstruction with Contrastive Signals for Balanced Visual Representation (CVPR 2026 · better on 9 of 9 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unified Generative and Discriminative Training for Multi-modal Large Language Models (NeurIPS 2024 · similarity 0.88)
- Aligning Generative Denoising with Discriminative Objectives Unleashes Diffusion for Visual Perception (ICLR 2025 · similarity 0.88)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.88)
- Learning Visual Generative Priors without Text (CVPR 2025 · similarity 0.88)
- Diffusion Feedback Helps CLIP See Better (ICLR 2025 · similarity 0.88)
- BiGR: Harnessing Binary Latent Codes for Image Generation and Improved Visual Representation Capabilities (ICLR 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.
