Comparing local version AF0.3 on Comfy to Playground?
Hi!
Wanted to say that AF0.3 is really something. Have it running in Comfy on a 3060/12GB, 50 steps in ~250s for 1024x1024.
Been playing with trying to coax AF0.3/Comfy into producing results similar to those from the instance currently running on Playground. This has proved difficult! Have tried with the online "orange juice" example (https://fal.ai/models/fal-ai/aura-flow?share=99d2cb8c-8393-4d58-b26c-dfbed2818c71), attempting to reproduce that outcome locally. Have iterated over parameters (steps, shift, CFG, sampler, scheduler) and haven't quite nailed it. (FYI, dpmpp2m/karras has come closest, but no cigar.)
We've also tried other experiments with custom prompts. Playground seems to adhere better to prompting, and we see this e.g., with the local 0.3 leaving out details that Playground will include (same seed, same prompt, same params). Also composition and anatomy are more faithful in Playground.
Perhaps the sampler/scheduler on Playground are custom or very specific, in a way that a local workflow can't reproduce.
Does anyone have specific insights into how to get better correlation? Or, is the Playground model simply different right now?
Thanks in advance for your insights!
the playground is running AF0.2 fyi
You can download the WebUI for it here: https://github.com/JeremyDemers/AuraFlow-WebUI