Player not loading? Watch on YouTube
This tutorial tests SANA image generation on a Windows laptop with an 8GB RTX 4060. The speaker compares SANA-Sprint 0.6B at two steps, SANA 1.6B at 20 steps, and a 3840×3840 model. The setup uses a Python virtual environment and installs Diffusers, Transformers, Accelerate and SentencePiece before loading the pipeline and saving an image.
The main lesson is memory management. The speaker reports that placing Sprint directly on the GPU took 12.81 seconds per image and peaked at 8.2GB. Replacing that placement with enable_model_cpu_offload() reduced the reported time to 4.40 seconds and peak memory to 4.97GB. He attributes the improvement to avoiding Windows memory spill across PCIe. A subsequent run produced 24 images in 126.24 seconds, with no internet connection after the download.
The comparison also shows a quality tradeoff. Sprint failed to spell a requested shop sign in the example; the 1.6B model rendered it correctly with the same prompt and seed, but took about 27 seconds per image. The spelling result comes from that specific prompt. For 4K output, CPU offload plus decoder tiling reportedly kept peak memory at 5.08GB and completed an image in 151.97 seconds.
For offline local AI use, the download is a practical constraint: the 0.6B pipeline occupies about 7.2GB, mostly for its Gemma text encoder. The first load also takes several minutes. These laptop measurements differ from the project's advertised H100 and RTX 4090 timings.
