CVPR 2026 | Diffusion开始瘦身：TinySR提速5.68倍，走向手机端部署
Ranking
Overall
73
Content
85
Popularity
44
Observed public metrics from 1 member.
Merged summary
TL;DR - TinySR is a lightweight, one-step diffusion-transformer system for real-world image super-resolution, designed around practical mobile deployment. It reports up to 5.68× faster inference than its TSD-SR teacher while reducing parameters by 83% and MACs by 84%, with competitive perceptual quality.
- Uses learnable probabilistic masks, Dynamic Inter-block Activation, and an Expansion-Corrosion strategy to identify and progressively prune recoverable transformer blocks.
- Compresses the VAE through channel pruning, attention removal, and depthwise separable convolutions, yielding about 10× VAE inference acceleration and 22× fewer VAE MACs.
- Removes low-value text- and time-conditioning modules for the specialized one-step super-resolution task, further reducing model size and latency with limited metric changes.
- Pre-caches stable modulation parameters to eliminate repeated inference-time computation and improve end-to-end deployment efficiency.
Sources (1)
CVPR 2026 | Diffusion开始瘦身：TinySR提速5.68倍，走向手机端部署
Public signals
Hugging Face upvotes 1
TL;DR - TinySR is a lightweight, one-step diffusion-transformer system for real-world image super-resolution, designed around practical mobile deployment. It reports up to 5.68× faster inference than its TSD-SR teacher while reducing parameters by 83% and MACs by 84%, with competitive perceptual quality.
- Uses learnable probabilistic masks, Dynamic Inter-block Activation, and an Expansion-Corrosion strategy to identify and progressively prune recoverable transformer blocks.
- Compresses the VAE through channel pruning, attention removal, and depthwise separable convolutions, yielding about 10× VAE inference acceleration and 22× fewer VAE MACs.
- Removes low-value text- and time-conditioning modules for the specialized one-step super-resolution task, further reducing model size and latency with limited metric changes.
- Pre-caches stable modulation parameters to eliminate repeated inference-time computation and improve end-to-end deployment efficiency.