Got a question for “pipeline_onnx_stable_diffusion.py” maybe someone could help what’s wrong and its origin.
Scenario:
I’m getting the latents using a callback for one specific prompt (i.e. car)(line 425), saving one of them to disk and recovering it to apply it for a next run (dog), using its “Arg:latents (np.ndarray, optional)”, but any time i submit that latent to the call with a new prompt(dog), the resultant image does not change it always keep giving the same output ( car), The image provided is the same one as the output of the first run (car), ignoring the prompt of the second one, does not matter using same scheduler or not, number of steps or prompt… (meaning saved latents works and it could be related to the processing of received latents?.. )
Any ideas on how to use those previous latents?
PD: I tried relocating Line 385 to line 389, to avoid applying the scheduler sigmas two times, resulting in a “blank” image…