TensorSharp adds Qwen image editing, with CUDA speed results
TensorSharp now supports image editing and with models. The comparison tested TensorSharp against with the same input image, prompt, , step count, CFG setting, and seed. The run used CUDA, a 544x1184 image size, and 4 steps.
On a server that was already running, TensorSharp finished the full warm request in 40.44 seconds, while took 48.16 seconds. Per sampling step, TensorSharp took 7.57 seconds and took 9.43 seconds. Text encoding was slower on TensorSharp at 7.45 seconds, compared with 4.47 seconds on .
Image encoding and decoding were faster on TensorSharp, at 0.54 seconds and 1.51 seconds, compared with 1.92 seconds and 2.57 seconds. TensorSharp’s first cold request took 54.11 seconds because it had extra setup work on a fresh server.
Key points
- TensorSharp now supports for image editing and generation.
- In the CUDA test, TensorSharp completed a warm request in 40.44 seconds.
- took 48.16 seconds under the same test setup.
- TensorSharp’s first cold request took 54.11 seconds because of extra setup work.
- Mac mini users should not treat CUDA results as Mac numbers.