Research map / Image generation / Super-resolution
Edge-SD-SR: Low Latency and Parameter Efficient On-device Super-Resolution with Stable Diffusion via Bidirectional Conditioning
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.
Edge-SD-SR compresses a Stable Diffusion super-resolution model to ~169M parameters and ~142 GFLOPs and runs it in one step on a phone, using bidirectional conditioning, joint LR-encoder/UNet training, and decoder finetuning.
Problem
Stable Diffusion-based super-resolution is too large and slow (many denoising steps, high-resolution tiling) to run on mobile devices.
Approach
Condition both the forward noising and the Gaussian noise variance on the LR image (bidirectional conditioning), so sampling starts from the LR-derived distribution and the UNet needs no explicit LR input; jointly train a lightweight LR encoder with the UNet, combine with scale distillation for one-step inference, and finetune a small decoder on the UNet output.
Key results
- Edge-SD-SR reaches FID 25.37 and LPIPS 0.249 on DIV2K validation, versus 30.45 FID for ResShift and 26.47 for LDM, with 169M parameters and 142 GFLOPs.
- Bidirectional conditioning with variance conditioning reduces FID from 46.46 (unidirectional-free baseline) to 25.37 in the ablation.
- The model upscales a 128x128 patch to 512x512 in about 38 ms on a Samsung S24 NPU with W8A16 quantization, and a 512x512 image to 2048x2048 in about 1.1 s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use bidirectional conditioning instead of unidirectional conditioning (ablation) | Edge-SD-SR · DIV2K Valid (RealESRGAN degradations) | FID: 46.46 -> 25. 37 | Helped |
| Use bidirectional conditioning instead of unidirectional conditioning (ablation) | Edge-SD-SR · DIV2K Valid (RealESRGAN degradations) | LPIPS: 0.468 -> 0.249 | Helped |
| Add variance conditioning to bidirectional conditioning (ablation) | Edge-SD-SR · DIV2K Valid (RealESRGAN degradations) | FID: 39.13 -> 25. 37 | Helped |
| Add variance conditioning to bidirectional conditioning (ablation) | Edge-SD-SR · DIV2K Valid (RealESRGAN degradations) | LPIPS: 0.313 -> 0.249 | Helped |
| Train the LR encoder jointly with the UNet instead of independently (ablation) | Edge-SD-SR · DIV2K Valid (RealESRGAN degradations) | FID: 63.03 -> 25. 37 | Helped |
| Train the LR encoder jointly with the UNet instead of independently (ablation) | Edge-SD-SR · DIV2K Valid (RealESRGAN degradations) | LPIPS: 0.549 -> 0.249 | Helped |
| Fine-tune the efficient decoder using the UNet output instead of independent VQAE training (ablation) | Edge-SD-SR · DIV2K Valid (RealESRGAN degradations) | FID: 52.78 -> 25. 37 | Helped |
| Fine-tune the efficient decoder using the UNet output instead of independent VQAE training (ablation) | Edge-SD-SR · DIV2K Valid (RealESRGAN degradations) | LPIPS: 0.401 -> 0.249 | Helped |
Surprising findings
- Training a small decoder independently of the diffusion model gives much worse results than finetuning it on the UNet output.
Limitations
- Trails YONOS-SR on several benchmarks, which uses a much larger model with one step.
- Ablation results are on a single synthetic DIV2K validation set.
- Compared baselines are not matched in size or step count, so the comparison is not fair by design, as the authors note.
Experimental setup
Datasets: DIV2K DIV8K Flickr2k OST FFHQ subset DIV2K validation (RealESRGAN degradations) RealSR DRealSR DPED-iphone
Benchmarks: DIV2K validation (RealESRGAN degradations) RealSR DRealSR DPED-iphone
Metrics: FID LPIPS PSNR SSIM MUSIQ
Baselines: LDM StableSR YONOS-SR DiffBIR SeeSR ResShift
Builds on: Stable Diffusion 1.5 YONOS-SR scale distillation UNet pruning recipe (BK-SDM-style) ResShift InDI RealESRGAN degradation
Closest papers by meaning
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.90)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.90)
- BiMaCoSR: Binary One-Step Diffusion Model Leveraging Flexible Matrix Compression for Real Super-Resolution (ICML 2025 · similarity 0.89)
- PocketSR: The Super-Resolution Expert in Your Pocket Mobiles (NeurIPS 2025 · similarity 0.89)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.89)
- Unleashing the Power of One-Step Diffusion based Image Super-Resolution via a Large-Scale Diffusion Discriminator (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.