vmobilis

@stduhpf, this can be fixed, maybe easily ([no, it's not that easy, thank you](https://github.com/leejet/stable-diffusion.cpp/pull/681), @stduhpf), since GGML / SD.CPP support F64 and I64 types: https://github.com/ggml-org/ggml/blob/ff9052988b76e137bcf92bb335733933ca196ac0/include/ggml.h#L379 The function https://github.com/leejet/stable-diffusion.cpp/blob/10c6501bd05a697e014f1bee3a84e5664290c489/model.cpp#L961 should be...

@alexdsh, have you tried to run it with `--ops-printf` option, maybe it will show something useful?

@alexdsh, you need to specify a folder where `stable-diffusion-xl-base-1.0-onnxstream` subfolder is placed in: --models-path **FOLDER** --xl Where SDXL 1.0 is located in **FOLDER**/stable-diffusion-xl-base-1.0-onnxstream. Or simply launch `sd.exe --xl` from FOLDER....

@vitoplantamura, hello, I slightly changed the sampling cycle: instead of using hw counter, I'm simply checking x_ptr, this approach removes 1 increment operation. The changes are visible here: https://github.com/vitoplantamura/OnnxStream/commit/523bc53cab6d780a5d2b03c72b9d378f6f37c714#diff-5cb305592bc405473a9934a751e61fa93aafa9d3b2f6bd67de62af51ab53e4d0R1346. Also,...

@vitoplantamura, I tested some more, it looks like some samplers do not work. Images Model Euler Ancestral Euler DPM++ (2M) DPM++ (2S) Ancestral SD 1.5,15 steps SDXL,15 steps SDXL Koala...

@vitoplantamura, tried dpm++2m (updated examples above), something is definitely wrong. Poor kittens look like they were washed.

@vitoplantamura, DPM++ (2S) Ancestral [works better](https://github.com/vitoplantamura/OnnxStream/pull/120#issuecomment-2833484459), but Turbo model still does not look like a photo. And since Euler Ancestral works well with Turbo, I think it's not the model.

@vitoplantamura, I think I found a workaround: https://github.com/vitoplantamura/OnnxStream/blob/d5374bf029f3bd0baf7036e3353eb9d87e1455a8/src/sd.cpp#L1420-L1430 The problem looks related to noise scheduler, by slightly lowering sigma values it's possible to get more acceptably looking results: I also...

@vitoplantamura, hello, I [tried](https://github.com/vitoplantamura/OnnxStream/pull/120/commits/ae01cec044a0ec96ae41e6e25a857ff634266cf1) DPM++ 3M SDE from [k-diffusion](https://github.com/crowsonkb/k-diffusion), it works, but it needs too many steps, so maybe it's not really useful. Images from dpm++3msde_a Model 5 steps 15...

@vitoplantamura, thank you, I suspected that assumption "who cares about difference in sqrt() and std::sqrt()" could be wrong... 😄 I think it should compile now.