Research map / Image generation / Super-resolution
Arbitrary-Scale Image Generation and Upsampling using Latent Diffusion Model and Implicit Neural Decoder
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a latent diffusion model with an implicit neural decoder (symmetric auto-decoder plus LIIF-style MLP) for arbitrary-scale image generation and super-resolution, trained with a two-stage alignment loss.
Problem
Existing generative and SR models work only at fixed integer scales, and diffusion-based arbitrary-scale SR (IDM) runs implicit decoding inside every denoising step, which is slow and memory-hungry while lacking cross-scale consistency.
Approach
Run the diffusion process in the latent space of a pre-trained auto-encoder and decode with an implicit neural decoder (auto-decoder convolutions followed by an LIIF MLP), so the MLP is used only once per sample; a reconstruction loss backpropagated through the frozen decoder aligns the latent diffusion with image space.
Key results
- On CelebA-HQ 5.3x SR, the model reaches 24.66 PSNR / 0.0455 LPIPS versus IDM's 23.34 / 0.0526 and LIIF's 27.52 / 0.1207 at the same scale.
- On LSUN Bedroom 128 resolution, FID drops to 7.20 from 10.15 for ScaleParty, with recall 38.26 vs 20.63.
- Inference is about 12.7 times faster than IDM at 8x (0.2568 vs 0.0202 FPS).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed arbitrary-scale generator instead of ScaleParty | LSUN Bedroom image generation · LSUN Bedroom 128, 160, and 192 | FID: 10.15 -> 7.20; 9.85 -> 7.43; 9.91 -> 7.73 | Helped |
| use the proposed method instead of LIIF | CelebA-HQ super-resolution · 5.3× | LPIPS: 0.1207 -> 0.0455 | Helped |
| use the proposed method instead of IDM | CelebA-HQ super-resolution · 5.3× | PSNR: 23.34 -> 24.66 | Helped |
| use the proposed method instead of IDM | LSUN Tower super-resolution · 16× | PSNR: 19.44 -> 21.24 | Helped |
| use the proposed method instead of IDM | LSUN Tower super-resolution · 16× | LPIPS: 0.2549 -> 0.1897 | Helped |
| use the proposed method instead of LIIF | DIV2K super-resolution · 12× and 17× | PSNR: 22.28 -> 22.73; 21.23 -> 21.83 | Helped |
| move implicit decoding to a single decoding stage rather than each diffusion step | IDM inference · CelebA-HQ 8× | FPS: 0.0202 -> 0.2568 | Helped |
| add the symmetric auto-encoder decoder before the MLP (ablation) | implicit neural decoder · Decoder reconstruction 16 →128 | PSNR: 22.96 -> 35.28 | Helped |
| add the symmetric auto-encoder decoder before the MLP (ablation) | implicit neural decoder · Decoder reconstruction 64 →256 | PSNR: 32.54 -> 35.93 | Helped |
| add the two-stage alignment process (ablation) | proposed super-resolution method · CelebA-HQ 5.3× | PSNR: 22.79 -> 24.66 | Helped |
| add the two-stage alignment process (ablation) | proposed super-resolution method · CelebA-HQ 5.3× | LPIPS: 0.0600 -> 0.0455 | Helped |
Surprising findings
- LIIF keeps higher PSNR than the proposed method at in-distribution scales on CelebA-HQ (e.g., 27.52 vs 24.66 at 5.3x), so the gains are mainly in perceptual metrics.
Limitations
- Still slower than GAN-based or regression-based SR models.
- PSNR at in-distribution scales trails LIIF.
Experimental setup
Datasets: FFHQ CelebA-HQ LSUN DIV2K Flickr2K
Benchmarks: FFHQ LSUN Bedroom CelebA-HQ LSUN Tower DIV2K
Metrics: FID Precision Recall SelfSSIM PSNR LPIPS SSIM FPS
Baselines: LIIF SR3 IDM CIPS ScaleParty MS-PE MSPIE EDSR ESRGAN SRFlow PULSE GLEAN
Builds on: Latent Diffusion Models LIIF IDM DDIM
Closest papers by meaning
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.90)
- Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution (NeurIPS 2026 · similarity 0.89)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.89)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.88)
- Unleashing the Power of One-Step Diffusion based Image Super-Resolution via a Large-Scale Diffusion Discriminator (NeurIPS 2025 · similarity 0.88)
- Adversarial Diffusion Compression for Real-World Image Super-Resolution (CVPR 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.