request if you do a v1.3
First of all amazing job. Beats klein 9b for my use case.
I have 2 requests only.
In the future could you please try to support 2mp generation atleast? Would be perfect.
And if possible i hope you can reduce the generation speed penalty a bit, or maybe figure out a way to do it in less steps.
Thanks!
Thank you, and both are on my mind for the next version. 2MP: the likeness drop above ~1536 is mostly because the base was trained around 1 to 2MP and the LoRA at 1024, so a higher-resolution training pass is the plan. Speed: it is slower than a plain generation because the source goes in as in-context reference tokens, which roughly doubles the sequence, so a real speedup likely needs a distilled edit-Turbo down the line. Noted and appreciated.
Thank you, and both are on my mind for the next version. 2MP: the likeness drop above ~1536 is mostly because the base was trained around 1 to 2MP and the LoRA at 1024, so a higher-resolution training pass is the plan. Speed: it is slower than a plain generation because the source goes in as in-context reference tokens, which roughly doubles the sequence, so a real speedup likely needs a distilled edit-Turbo down the line. Noted and appreciated.
I noticed on the workflow it says trained at 768, is that just an old note and i can increase grounding res to 1024?
Grounding res is not the training res. The Lora is trained at 1024 resolution, with the grounding res shuffled between 384 and 768. You can definitely try to go higher res and also use higher grounding res if you want to, though resolutions outside of the trained range are less reliable. :)