Research map / Image generation / Autoregressive and latent generation
DiT-IC: Aligned Diffusion Transformer for Efficient Image Compression
Junqi Shi, Ming Lu, Xingchen Li, Anle Ke, Ruiqi Zhang, Zhan Ma
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 alignment strategies (variance-guided flow, self-distillation, latent condition) to DiT baseline (ablation) | DiT-IC · Avg BD-rate LPIPS | BD-rate (%): -65.00 -> -83.65 | Improved |
| add alignment strategies to DiT baseline (ablation) | DiT-IC · Avg BD-rate DISTS | BD-rate (%): -67.19 -> -87.88 | Improved |
| initialize from pretrained SANA instead of training DiT from scratch (ablation) | DiT-IC · ablation | BD-rate DISTS (%): 32.45% -> 0% | Improved |
| use LoRA rank 32/64 instead of full finetuning (ablation) | DiT-IC · ablation | BD-rate DISTS (%): 8.05% -> 0% | Improved |
| use LoRA rank 32/64 instead of rank 16/16 (ablation) | DiT-IC · ablation | BD-rate DISTS (%): 13.92% -> 0% | Improved |
| add adversarial loss Ladv (ablation) | DiT-IC · ablation | BD-rate PSNR (%): -37.10% -> 0% | Hurt |
| add adversarial loss Ladv (ablation) | DiT-IC · ablation | BD-rate DISTS (%): -1.80% -> 0% | Improved |
| add DISTS term to distortion loss (ablation) | DiT-IC · ablation | BD-rate DISTS (%): 5.69% -> 0% | Improved |
| add DISTS term to distortion loss (ablation) | DiT-IC · ablation | BD-rate LPIPS (%): -1.83% -> 0% | Hurt |
| use DiT in 32x latent space instead of U-Net diffusion (StableCodec) | DiT-IC · 1024x1024 FP32 | diffusion latency (s): 0.11 -> 0.055 | Improved |
| use DiT-IC at 4096x4096 | DiT-IC · 4096x4096 FP32 | total latency (s): 85 -> 3.3 | Improved |
| use DiT-IC VAE vs SD-VAE at 1024x1024 | DiT-IC · 1024x1024 FP32 | VAE latency (s): 0.19 -> 0.21 | Hurt |
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.
- StableCodec — StableCodec: Taming One-Step Diffusion for Extreme Image Compression (ICCV 2025) · better on 7 of 7 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
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.92)
- SD-DiT: Unleashing the Power of Self-supervised Discrimination in Diffusion Transformer (CVPR 2024 · similarity 0.90)
- UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction (NeurIPS 2026 · similarity 0.90)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.90)
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.90)
- DiT4Edit: Diffusion Transformer for Image Editing (AAAI 2025 · 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.
