Research map / Image generation / Autoregressive and latent generation
One Model, Many Budgets: Elastic Latent Interfaces for Diffusion Transformers
Moayed Haji-Ali, Willi Menapace, Ivan Skorokhodov, Dogyun Park, Anil Kag, Michael Vasilkovsky, Sergey Tulyakov, Vicente Ordonez, Aliaksandr Siarohin
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 ELIT latent interface (Read/Write layers) | DiT-XL · ImageNet-1K 512px | FID50K: 18.8 -> 11.1 | Improved |
| add multi-budget tail-drop training to ELIT | ELIT-DiT · ImageNet-1K 512px | FID50K: 11.1 -> 10.1 | Improved |
| add multi-budget training to ELIT | ELIT-UViT · ImageNet-1K 256px (with 0.25 CFG) | FID50K: 3.8 -> 3.7 | No clear effect |
| add ELIT to DiT-XL for video | DiT-XL · Kinetics-700 256px | FVD10K: 135.9 -> 116.5 | Improved |
| use tail dropping instead of random token dropping (ablation) | ELIT-DiT · ImageNet-1K 256px (25% tokens) | FID10K: 38.6 -> 36.3 | Improved |
| use ELIT tail drop instead of multi-patch-size training for variable budget (ablation) | DiT-B/2 · ImageNet-1K 256px (25% tokens) | FID10K: 85.25 -> 36.3 | Improved |
| group size 4x4 vs 1x1 for Read/Write (ablation) | ELIT-DiT · ImageNet-1K 256px | FID10K: 29.94 -> 26.53 | Improved |
| use a single group spanning the whole image (16x16) instead of 4x4 (ablation) | ELIT-DiT · ImageNet-1K 256px | FID10K: 26.53 -> 30.03 | Hurt |
| add spatial head and tail blocks around the latent core (ablation) | ELIT-DiT · DiT-B/2 ImageNet-1K 256px | FID10K: 33.84 -> 26.53 | Improved |
| increase head/tail to 5 each with only 2 core blocks (ablation) | ELIT-DiT · DiT-B/2 ImageNet-1K 256px | FID10K: 25.34 -> 26.95 | Hurt |
| variable batch size vs constant batch size (ablation) | ELIT-DiT · ImageNet-1K 256px | FID10K: 26.15 vs 26.65 | No clear effect |
| reduce inference tokens in ELIT-Qwen-Image from 100% to 12.5% (ablation) | ELIT-Qwen-Image · DPG-Bench average | score: 90.45 -> 88.02 | Mixed |
Closest papers by meaning
- LEDiT: Your Length-Extrapolatable Diffusion Transformer without Positional Encoding (NeurIPS 2025 · similarity 0.89)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.89)
- Rethinking Cross-Layer Information Routing in Diffusion Transformer (NeurIPS 2026 · similarity 0.88)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.88)
- LaVin-DiT: Large Vision Diffusion Transformer (CVPR 2025 · similarity 0.88)
- UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction (NeurIPS 2026 · 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.
