Research map / Image generation / Autoregressive and latent generation
Toward Diffusible High-Dimensional Latent Spaces: A Frequency Perspective
Bolin Lai, XuDong Wang, Saketh Rambhatla, James M. Rehg, Zsolt Kira, Rohit Girdhar, Ishan Misra
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 FreqWarm frequency warm-up to diffusion training | USiT-H on Wan2.2-AE · ImageNet 512x512 w/o CFG | gFID: 43.67 -> 29.56 | Improved |
| add FreqWarm frequency warm-up | USiT-H on LTX-AE · ImageNet 512x512 w/o CFG | gFID: 24.18 -> 18.05 | Improved |
| add FreqWarm frequency warm-up | USiT-H on DC-AE-f32c128 · ImageNet 512x512 w/o CFG | gFID: 13.84 -> 9.42 | Improved |
| add FreqWarm frequency warm-up | USiT-2B on DC-AE-f64c128 · ImageNet 512x512 w/o CFG | gFID: 5.67 -> 4.77 | Improved |
| add FreqWarm frequency warm-up | DiT-XL on DC-AE-f32c128 · ImageNet 512x512 with CFG | gFID: 3.11 -> 2.87 | Improved |
| add FreqWarm frequency warm-up | DiT, UViT, USiT on DC-AE-f32c128 · ImageNet 256x256 | gFID: 26.30 -> 17.89 (DiT) | Improved |
| increase latent channels from 16 to 512 with FreqWarm (ablation) | DC-AE f32 · ImageNet 512x512 | gFID gain: 0.02 at c16 -> 12.18 at c512 | Improved |
| apply FreqWarm to low-channel autoencoder (ablation) | DC-AE-f32c16 · ImageNet 512x512 | gFID: 12.59 -> 12.57 | No clear effect |
| use FreqWarm threshold 0.2 instead of 0.05 (ablation) | DC-AE-f32c128 + USiT-H · ImageNet 512x512 | gFID: 23.11 -> 9.42 | Improved |
| use FreqWarm threshold 0.2 instead of 0.4 or 0.6 (ablation) | DC-AE-f32c128 + USiT-H · ImageNet 512x512 | gFID: 12.88 / 13.24 -> 9.42 | Improved |
| add FreqWarm to 512-channel autoencoder (ablation) | DC-AE-f32c512 · ImageNet 512x512 | gFID: 54.84 -> 42.66 | Improved |
Closest papers by meaning
- Improving the Diffusability of Autoencoders (ICML 2025 · similarity 0.92)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.91)
- FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 2025 · similarity 0.91)
- UltraDiff:Transferring High-Fidelity Priors to Compressed Latent Spaces for High-resolution Image Generation (NeurIPS 2026 · similarity 0.90)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.90)
- Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think (NeurIPS 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.
