When generating images with Stable Diffusion, you&#39;d want an upscale converter to further improve quality, right? Let&#39;s use the standard one provided. First, let&#39;s produce an image as usual. This time, I&#39;ll try it with my favorite supercar.The GPU is an RTX 4090, which is quite fast.The image size is the standard 512x512. Let&#39;s try with 20 steps for the sample. RTX4090 / 512x512 / 20step / 2.9s<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/20/00031-2629032906_normal_29s.png"/> with 100 steps, RTX4090 / 512x512 / 100step / 11.6s<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/20/512_100step_116s.png"/> It&#39;s a bit detailed, but the composition seems off... Next, let&#39;s try upscaling. With diffusion models, even using the same seed, starting with a larger image size results in a different picture. To upscale while maintaining the composition, we use an upscale converter. I&#39;ve tried upscaling to 1024 pixels. For the sampler, I chose latent. Let&#39;s set the base model to 20 steps and the sampler to 20 steps as well. RTX4090 / 512x512 to 1024x1024 / base20step, hires20step / latent / denoising 0.7 / 11.6s<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/20/00026-2629032906_latent_116s.png"/> It looks good, but the logo has doubled, and the right rear part has swelled a bit... RTX4090 / 512x512 to 1024x1024 / base20step, hires20step / DATx2 / denoising 0.7 / 55.5s<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/20/00039-2629032906_222222.png"/> The details are quite different. There was also a 5x difference in speed. It seems there&#39;s still room for improvement in how the model renders details. The upscaling went well, too.

Stable Diffusion + Standard Upscale Converter / RTX 4090 Benchmark

Yuichiro Minato