Research map / Image generation / Autoregressive and latent generation
Variable-Length Tokenization via Learnable Global Merging for Diffusion Transformers
Dong Hoon Lee, Seunghoon Hong
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 |
|---|---|---|---|
| replace nested dropout with learnable global merging for length modulation (ablation) | variable-length tokenizer + LightningDiT-B (joint) · ImageNet 256x256, 32 tokens | gFID: 5.21 -> 3.19 | Improved |
| replace nested dropout with global merging (ablation) | variable-length tokenizer · ImageNet 256x256 cross-length | CKNNA: 0.34 -> 0.76 average | Improved |
| train one jointly across lengths vs separate models per length (nested dropout) (ablation) | nested dropout latents · ImageNet 256x256, 128 tokens | gFID: 2.94 -> 4.69 | Hurt |
| joint instead of separate training with learnable global merging (ablation) | global merging latents · ImageNet 256x256, 128 tokens | gFID: 2.48 -> 2.79 | Hurt |
| add length-specific LoRA post-training (10 epochs) (ablation) | jointly trained DiT on merged latents · ImageNet 256x256, 128 tokens | gFID: 2.79 -> 2.52 | Improved |
| add LoRA at 32 tokens (ablation) | jointly trained DiT on merged latents · ImageNet 256x256, 32 tokens | gFID: 3.19 -> 3.04 | Improved |
| add global (data-independent) merging with proportional attention and merged positional embeddings over ToMe-style merging (ablation) | DiT-B · ImageNet 256x256, 128 tokens | gFID: 4.89 -> 4.27 | Improved |
| add learnable embeddings to global merging (ablation) | DiT-B · ImageNet 256x256, 128 tokens | gFID: 4.27 -> 3.21 | Improved |
| add alignment loss to learnable global merging (ablation) | DiT-B · ImageNet 256x256, 128 tokens | gFID: 3.21 -> 2.79 | Improved |
| use separate positional embeddings per token count instead of merged ones (ablation) | DiT · ImageNet 256x256 | gFID | No clear effect |
| use data-dependent ToMe merging instead of global merging for reconstruction (ablation) | tokenizer · ImageNet 256x256 reconstruction | rFID/PSNR | No clear effect |
| more diffusion sampling steps (25 to 100) for our method (ablation) | Ours · ImageNet 256x256, 32 tokens | gFID: 2.89 -> 2.54 | Improved |
Closest papers by meaning
- ToMA: Token Merge with Attention for Diffusion Models (ICML 2025 · similarity 0.91)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.91)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.90)
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.89)
- Which Tokens to Merge? Diffusion Dynamics for Efficient Image Generation (NeurIPS 2026 · similarity 0.89)
- End-to-End Training for Unified Tokenization and Latent Denoising (NeurIPS 2026 · similarity 0.89)
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.
