Could you provide a complete inference example?
@ostris @0-0 @0-0-0-0-0-0-0-0-0-0-0-0-0-0-0-0-0-0-0-0-0a @0-0-WasAlreadyTaken @0-002 I tried the open_flux_pipeline.py you provided, but the inference results were not as good as Flux Schnell. Could you give a full inference example?
You need to load openflux1-v0.1.0-fast-lora.safetensors before inference, but I also don't understand why the results are not good when running the non-accelerated version directly.
It should be fine to run it without the accelerator, at 20 steps. Do you have both kinds of guidance set up in your inference pipeline? The real CFG and the "fake" guidance. As far as I understand, with OpenFlux both types of guidance remain active during inference, affecting the results, whether or not one actually sets up a way to alter their parameter values or not. Or is that not so? (A question for Ostris, or for anyone else with greater insight/experience than me re: technical features/dynamics.)