Research map / Image generation / Autoregressive and latent generation
Concept-Guided Tokenization: Closing the Gap Between Reconstruction and Generation
Yunqiao Yang, Haokun Lin, Guanzhong Wu, Ying Wei
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 concept guidance (SAE concept-index loss) to the tokenizer (ablation) | ConceptTok-B-B-64 baseline · ImageNet C2I (LlamaGen-B) | gFID: 7.31 -> 6.51 | Improved |
| add text conditioning in the tokenizer encoder (ablation) | ConceptTok-B-B-64 baseline · ImageNet C2I (LlamaGen-B) | gFID: 7.31 -> 5.31 | Improved |
| combine text conditioning and concept guidance (ablation) | ConceptTok-B-B-64 · ImageNet C2I (LlamaGen-B) | gFID: 5.31 -> 4.13 | Improved |
| use concept guidance instead of holistic cosine alignment (ablation) | ConceptTok-B-B-64 with text conditioning · ImageNet C2I (LlamaGen-B) | gFID: 4.84 -> 4.13 | Improved |
| use concept guidance instead of cosine alignment (ablation) | ConceptTok-B-B-64 · ImageNet CKNNA | CKNNA: 0.39 -> 0.48 | Improved |
| use concept-space alignment instead of dense feature alignment (REPA) for the diffusion generator | SiT-L/2 · ImageNet C2I 256x256, 400K iterations | gFID: 9.7 -> 8.7 | Improved |
| increase number of latent tokens from 32 to 128 (ablation) | ConceptTok-B-L · ImageNet C2I (LlamaGen-B) | gFID: 6.19 -> 2.97 | Improved |
| increase tokenizer size from S-S to B-B (64 tokens) (ablation) | ConceptTok · ImageNet C2I (LlamaGen-B) | gFID: 7.13 -> 4.13 | Improved |
| increase concept loss weight lambda from 0.1 to 0.5 (ablation) | ConceptTok-B-B-64 · ImageNet C2I (LlamaGen-B) | gFID: 4.13 -> 4.84 | Hurt |
| add concept guidance with lambda 0.1 versus lambda 0 (ablation) | ConceptTok-B-B-64 · ImageNet val reconstruction | rFID: 4.66 -> 3.84 | Improved |
| use CLIP/B instead of SigLIP/B as the pre-trained model for the SAE (ablation) | ConceptTok-S-S-64 · ImageNet C2I (LlamaGen-B) | gFID: 7.13 -> 6.80 | No clear effect |
| replace class-template captions with descriptive captions in training (ablation) | ConceptTok-B-L-128 · COCO-30k T2I | gFID: 10.73 -> 10.16 | Improved |
Closest papers by meaning
- Language-Guided Image Tokenization for Generation (CVPR 2025 · similarity 0.92)
- Image Understanding Makes for A Good Tokenizer for Image Generation (NeurIPS 2024 · similarity 0.91)
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.90)
- End-to-End Vision Tokenizer Tuning (NeurIPS 2025 · similarity 0.90)
- UniTok: a Unified Tokenizer for Visual Generation and Understanding (NeurIPS 2025 · similarity 0.90)
- A More Word-like Image Tokenization for MLLMs (CVPR 2026 · similarity 0.90)
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.
