Research map / Image generation / Autoregressive and latent generation
DC-Gen: Post-Training Diffusion Acceleration with Deeply Compressed Latent Space
Wenkun He, Yuchao Gu, Junyu Chen, Junyi Wu, Wenhang Ge, Dongyun Zou, Yujun Lin, Zhekai Zhang, Haocheng Xi, Muyang Li, Ligeng Zhu, Jincheng YU and 4 more
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 embedding alignment before fine-tuning, vs direct fine-tuning, when moving to DC-AE-f32 latents (ablation) | Wan2.1-T2V-1.3B · video generation training stability | training stability / visual quality | Improved |
| add embedding alignment vs direct fine-tuning (ablation) | FLUX.1-Krea · text-to-image visual quality | image detail | Improved |
| add patch embedder alignment (ablation) | DC-Gen · FLUX.1-Krea / Wan2.1-1.3B | quality and stability | Improved |
| add output layer alignment (ablation) | DC-Gen · end-to-end fine-tuning convergence | steps to converge: at least double steps without it | Improved |
| use embedding alignment instead of layer-wise LR decay (ablation) | FLUX.1-Krea · MJHQ-30K 512 | FID: 17.83 -> 13.30 | Improved |
| use embedding alignment instead of gradual LoRA (ablation) | FLUX.1-Krea · MJHQ-30K 512 | FID: 15.90 -> 13.30 | Improved |
| use embedding alignment instead of autoencoder latent space alignment (ablation) | FLUX.1-Krea-12B · MJHQ-30K 512 | gFID: 15.27 -> 13.30 | Improved |
| use LoRA instead of full fine-tuning in end-to-end stage (ablation) | FLUX.1-Krea · MJHQ-30K 512 | CLIP Score: 26.85 -> 27.38 | Improved |
| use LoRA instead of full fine-tuning in end-to-end stage (ablation) | Wan2.1-T2V-1.3B · VBench 480x832 | Overall: 79.81 -> 83.48 | Improved |
| switch to deeply compressed latent (f32) via DC-Gen | FLUX.1-Krea-12B · MJHQ-30K 1024 GenEval | GenEval and latency: GenEval 0.69 -> 0.71; latency 4.06 -> 0.88 s | Improved |
| switch to deeply compressed latent via DC-Gen | Wan2.1-T2V-14B · VBench 720x1280 | Overall / latency: 83.73 -> 83.97; 27.52 -> 3.58 min | Improved |
| switch to deeply compressed latent via DC-Gen | Qwen-Image-Edit · GEdit 1K English | Overall (O): 7.23 -> 7.20 (tuned vs DC-Gen); latency 57.63 -> 11.96 s | No clear effect |
Closest papers by meaning
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.91)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.90)
- DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence (NeurIPS 2026 · similarity 0.90)
- Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.89)
- Improving the Diffusability of Autoencoders (ICML 2025 · similarity 0.88)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 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.
