Research map / Image generation / Autoregressive and latent generation
LUA: Latent Upscaling Adapter for Diffusion-Based Image Synthesis
Aleksandr Razin, Kazantsev Danil, Ilya Makarov
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 Stage III edge-aware pixel refinement (ablation) | LUA curriculum · OpenImages val (x2) | PSNR: 28.96 -> 32.54 | Improved |
| add Stage II joint latent-pixel consistency (ablation) | LUA curriculum · OpenImages val (x2) | PSNR: 31.05 -> 32.54 | Improved |
| add Stage I latent structural alignment (ablation) | LUA curriculum · OpenImages val (x2) | PSNR: 31.60 -> 32.54 | Improved |
| add pixel-domain stages to latent l1-only training (ablation) | LUA curriculum · OpenImages val (x4) | LPIPS: 0.236 -> 0.184 | Improved |
| use joint multi-head instead of separate x2 specialist (ablation) | LUA · OpenImages val (x2) | PSNR: 31.92 -> 32.54 | Improved |
| use joint multi-head instead of separate x4 specialist (ablation) | LUA · OpenImages val (x4) | PSNR: 27.71 -> 27.94 | Improved |
| use pixel-shuffle heads instead of implicit LIIF upsampler (ablation) | LUA · OpenImages val (x2) | PSNR: 29.10 -> 32.54 | Improved |
| replace pixel-space SwinIR with latent-space LUA | SDXL 1024 base · OpenImages 2048x2048 | FID: 183.16 -> 180.80 | Improved |
| replace direct high-res sampling with LUA upscaling | SDXL · OpenImages 4096x4096 | FID: 280.42 -> 176.90 | Improved |
| replace direct SDXL sampling with LUA at 1K | SDXL · OpenImages 1024x1024 | FID: 194.53 -> 209.80 | Hurt |
| replace SwinIR with LUA at 1K | SDXL · OpenImages 1024x1024 | pFID: 204.23 -> 191.75 | Improved |
| increase fine-tuning pairs for cross-VAE transfer from 10k to 500k | LUA on SDXL · OpenImages 2048x2048 | FID: 203.1 -> 180.80 | Improved |
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.
- DemoFusion — DemoFusion: Democratising High-Resolution Image Generation With No $$$ (CVPR 2024) · better on 2 of 3 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
- DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.90)
- ScaleDiff: Higher-Resolution Image Synthesis via Efficient and Model-Agnostic Diffusion (NeurIPS 2025 · similarity 0.89)
- SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis (ICLR 2024 · similarity 0.89)
- When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution (NeurIPS 2026 · similarity 0.89)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.89)
- PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · 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.
