Research map / Vision-language and multimodal / Multimodal LLMs
TokLIP: Marry Visual Tokens to CLIP for Multimodal Comprehension and Generation
Haokun Lin, Teng Wang, Yixiao Ge, Yuying Ge, Zhichao Lu, Ying Wei, Qingfu Zhang, Zhenan Sun, 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 |
|---|---|---|---|
| add [CLS] token distillation from teacher (ablation) | TokLIP with contrastive loss · CC3M zero-shot ImageNet | Top-1 accuracy: 32.93 -> 52.46 | Improved |
| initialize vision encoder with pretrained weights (ablation) | TokLIP · CC3M zero-shot ImageNet | Top-1 accuracy: 15.73 -> 32.93 | Improved |
| use MLP projection instead of codebook mapping (ablation) | TokLIP · CC3M zero-shot ImageNet | Top-1 accuracy: 35.17 -> 52.46 | Improved |
| use MLP instead of soft MLP projection (ablation) | TokLIP · CC3M zero-shot ImageNet | Top-1 accuracy: 51.53 -> 52.46 | Improved |
| distill all tokens instead of [CLS] only (ablation) | causal TokLIP · CC3M zero-shot ImageNet | Top-1 accuracy: 52.46 -> 51.04 | Hurt |
| distill all tokens instead of single token (ablation) | bidirectional TokLIP · CC3M zero-shot ImageNet | IN Top1: 56.32 -> 59.47 | Improved |
| use bidirectional instead of causal attention (ablation) | TokLIP · CC3M comprehension | IN Top1: 52.46 -> 56.32 | Improved |
| use bidirectional instead of causal attention (ablation) | TokLIP · ImageNet generation (LlamaGen-B) | FID: 15.05 -> 214.36 | Hurt |
| fuse VQGAN and semantic features by concatenation instead of sum (ablation) | TokLIP generation · ImageNet generation (LlamaGen-B, 384, 10K samples) | FID: 16.37 -> 14.93 | Improved |
| add semantic features with learnable weighted sum (ablation) | VQGAN generation · ImageNet generation (LlamaGen-B, 384, 10K samples) | FID: 17.31 -> 15.46 | Improved |
| replace VQGAN tokenizer with TokLIP | LlamaGen GPT-B 384 · ImageNet generation (50 epochs) | FID: 14.48 -> 12.37 | Improved |
| replace VQGAN tokenizer with TokLIP | LlamaGen GPT-B 256 · ImageNet generation (50 epochs) | FID: 13.21 -> 13.12 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports TokLIP as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- SemHiTok: A Unified Image Tokenizer via Semantic-Guided Hierarchical Codebook for Multimodal Understanding and Generation (ICLR 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- QLIP — QLIP: A Dynamic Quadtree Vision Prior Enhances MLLM Performance Without Retraining (ICLR 2026) · better on 4 of 4 shared results
- TokenFlow-XL — TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025) · better on 1 of 1 shared result
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
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.88)
- ViCToR: Improving Visual Comprehension via Token Reconstruction for Pretraining LMMs (AAAI 2026 · similarity 0.88)
- UniTok: a Unified Tokenizer for Visual Generation and Understanding (NeurIPS 2025 · similarity 0.87)
- VITRIX-CLIPIN: Enhancing Fine-Grained Visual Understanding in CLIP via Instruction-Editing Data and Long Captions (NeurIPS 2025 · similarity 0.87)
- From Pixels to Tokens: Byte-Pair Encoding on Quantized Visual Modalities (ICLR 2025 · similarity 0.87)
- Auto-Encoding Morph-Tokens for Multimodal LLM (ICML 2024 · 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.
