OSEDiff
One-step diffusion super-resolution: diffusion-quality upscaling at a tenth of a second per image.
About
OSEDiff tackles the main complaint about diffusion-based super-resolution — speed. Diffusion SR models like SeeSR need 50 sampling steps and several seconds per image; OSEDiff distills the process into a single diffusion step using LoRA fine-tuning on Stable Diffusion 2.1-base, a degradation-aware prompt extractor, and a variational score distillation loss. The paper reports inference in about 0.11 seconds with image quality comparable to multi-step methods, at roughly 10x the speed of SeeSR.
The repo's published evals show it beating GAN baselines like Real-ESRGAN on perceptual metrics (MUSIQ, CLIPIQA) across DRealSR and RealSR sets while keeping only 8.5M trainable parameters. It was accepted at NeurIPS 2024 and the code is Apache 2.0, so it's a practical option for anyone who wants diffusion-level detail without a GPU farm — though as with all generative SR, it can hallucinate texture that wasn't in the original.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (arXiv 2406.08177): 1 diffusion step, 0.11s inference, 2,265G MACs vs SeeSR's 50 steps / 4.30s / 65,857G MACs
- Paper-reported: DRealSR MUSIQ 64.66 and CLIPIQA 0.6963, ahead of Real-ESRGAN (54.18 / 0.4422) and BSRGAN
- No independent third-party benchmarks found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: peer-reviewed NeurIPS 2024 paper (arXiv 2406.08177) with DRealSR/RealSR/DIV2K evals; Apache-2.0 repo with published metric tables. Strengths: dramatic, well-documented speed advantage; fully open. Docked for: (1) all benchmarks are author-reported with no independent replication found; (2) generative SR can hallucinate detail — a fidelity caveat, not a defect. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · NeurIPS 2024 paper arXiv 2406.08177; 1 diffusion step, 0.11s inference vs SeeSR 50 steps/4.30s; DRealSR MUSIQ 64.66 / CLIPIQA 0.6963 ahead of Real-ESRGAN
- · Apache-2.0 repo with published metric tables; only 8.5M trainable params
- · All benchmarks author-reported, no independent replication; generative SR can hallucinate detail