Research map / Image generation / Autoregressive and latent generation
GENERATING HIGH FIDELITY IMAGES WITH SUBSCALE PIXEL NETWORKS AND MULTIDIMENSIONAL UPSCALING
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use subscale ordering with 16x16 slices instead of the decoder baseline (ablation) | SPN hybrid decoder · ImageNet 32x32 | NLL (bits/dim): 3.79 -> 3.85 | Hurt |
| Use subscale ordering with 8x8 slices instead of the decoder baseline (ablation) | SPN hybrid decoder · ImageNet 32x32 | NLL (bits/dim): 3.79 -> 3.91 | Hurt |
| Use SPN subscale ordering rather than the decoder baseline (ablation) | SPN · ImageNet 64x64 | NLL (bits/dim): 3.52 -> 3.53 | Hurt |
| Add depth upscaling (ablation) | SPN · ImageNet 64x64 | NLL (bits/dim): 3.53 -> 3.53 | No clear effect |
| Use SPN instead of Glow | ImageNet 64x64 (5bit) · ImageNet 64x64 (5bit) | NLL (bits/dim): 1.76 -> 1.41 | Helped |
| Use SPN instead of Parallel Multiscale | ImageNet 128x128 · ImageNet 128x128 | NLL (bits/dim): 3.55 -> 3.08 | Helped |
| Add depth upscaling (ablation) | SPN · ImageNet 128x128 | NLL (bits/dim): 3.08 -> 3.08 | No clear effect |
| Add depth upscaling (ablation) | SPN · ImageNet 256x256 | NLL (bits/dim): 2.97 -> 3.01 | Hurt |
| Use SPN instead of Glow | CelebA-HQ 256x256 (5bit) · CelebA-HQ 256x256 (5bit) | NLL (bits/dim): 1.03 -> 0.61 | Helped |
| Apply multidimensional upscaling rather than depth upscaling alone | SPN samples · ImageNet 128x128 | sample fidelity | Helped |
| Use depth upscaling | SPN samples · ImageNet 128x128 | semantic coherence | Helped |
Closest papers by meaning
- PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · similarity 0.89)
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.88)
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · similarity 0.88)
- UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024 · similarity 0.87)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.87)
- PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining (NeurIPS 2026 · similarity 0.87)
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.