What sampling methods are available?
I have been struggling with samplers methods with ONNX, I can optimize models using Olive in SD 1.5 only though, and I only have few samplers available not all of them like without Olive.
Is that the case with this optimized model? and do you know how to get around this issue?
Support for Olive optimized models is extremely scarce right now unfortunately. Without knowing all the details, I'd say whatever doesn't work for SD1.5 won't work for SDXL either. What are you using to run your models though?
I am using Automatic1111 SD webui DirectML by lshqqytiger
Well, I optimized safetensors models myself using the A1111 +Olive tab, and can only manage to do with SD1.5 models, no SDXL so far would be successful.
The sampling methods are only few available, unlike th non-optimized versions with a long list, for example I can't find this SM at all DPM++ SDE Karras.
I am not sure if this is limitation because of Olive or is there something than needs to be done to unlock more sampling methods!
Would be really interested to understand more about this
Ah, lshqqytiger was the one who asked me to upload this model in the first place. I believe he is currently not working on supporting olive models in his own webui anymore, but instead trying to bring support to a similar webui called SD.Next. According to this comment.
However, I haven't been able to get SDXL Olive models to work at all with SD.Next yet. Though their wiki says it should work.
As for the samplers, SD.Next on the Olive branch is also missing quite a few of them. No idea what that's about, unfortunately. If you really need all the samplers, I suggest switching to Linux + ROCm for the time being, until Olive support becomes a little more widespread. I doubt there is any "hard limit" that prevents the samplers from working with Olive, it just needs to be implemented by someone.
Amazing, Thank you for clarifying that. I think it's related to the limited support to Olive/ONNX for AMD at the moment.
Also, maybe it's worth trying SD.Next but not expecting full support either.
I'll try Linux ROCm for now