Research map / Image generation / Autoregressive and latent generation
LacTokGen: Latent Consistency Tokenizer for 1024-pixel Image Generation by 256 Tokens
Qingsong Xie, Luyuan Zhang, Zhao Zhang, Siyuan Li, Zhe Huang, Zhenyu Yang, Haonan Lu
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 pixel reconstruction loss with consistency acceleration (LacTok-H) over VQ-LADD with diffusion loss only (ablation) | LacTok · MJHQ-5K 1024px reconstruction | rFID: 12.72 -> 11.34 | Improved |
| swap LlamaGen decoder for latent diffusion decoder trained with diffusion loss only (ablation) | LlamaGen tokenizer · MJHQ-5K 1024px reconstruction | PSNR: 19.24 -> 16.53 | Hurt |
| add HyperSD or TLCM acceleration without pixel loss to VQ-LADD (ablation) | VQ-LADD · MJHQ-5K 1024px reconstruction | rFID: 12.72 -> 14.71 (HyperSD), 14.40 (TLCM) | Hurt |
| use TLCM (two-step) instead of HyperSD (one-step) in the decoder (ablation) | LacTok with constructed data · FLUX-5K 1024px reconstruction | rFID: 12.18 -> 11.04 | Improved |
| train tokenizer on constructed FLUX-synthesized data instead of ImageNet (ablation) | LacTok-H · MSCOCO-2017 5K reconstruction | rFID: 10.80 -> 16.93 | Hurt |
| train tokenizer on constructed data (ablation) | LacTok-H · FLUX-5K reconstruction | rFID: 12.45 -> 12.18 | Improved |
| increase token count from 192 to 324 (ablation) | LacTok-H · MJHQ-5K 1024px reconstruction | rFID: 12.35 -> 11.01 | Improved |
| train autoregressive model on constructed high-quality data (ablation) | LacTokGen · GenEval | overall score: 0.71 -> 0.72 | Improved |
| use TLCM instead of HyperSD in LacTokGen (ablation) | LacTokGen · GenEval | overall score: 0.72 -> 0.73 | Improved |
| increase CFG scale above 2 (ablation) | LacTokGen-L* · MSCOCO-2017 5K generation | HPSv2: 0.303 (1.5) -> 0.304 (2, 3, 7) | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- HART — HART: Efficient Visual Generation with Hybrid Autoregressive Transformer (ICLR 2025) · better on 2 of 2 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
- An Image is Worth 32 Tokens for Reconstruction and Generation (NeurIPS 2024 · similarity 0.90)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.89)
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.89)
- FlexTok: Resampling Images into 1D Token Sequences of Flexible Length (ICML 2025 · similarity 0.88)
- ImageFolder: Autoregressive Image Generation with Folded Tokens (ICLR 2025 · similarity 0.88)
- Language-Guided Image Tokenization for Generation (CVPR 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.
