#fluxdepth#lora - still no result
logs like - something is changing but not in reality
. first tests was 64 samples with LR 1e-4 and 8k/3.5k steps rank64/a64
_
now 693 samples and 5e-5 \ -gradient_accumulation_steps 2 \ rank64/a64 (god help me plz)
testing my VHS-EVCORE lora based on my VHS stills (2014) year/
(made on jvc gr-ax66 vhs camera)
dataset contains 244 pictures. (without tape glitches)
my vhs stills archive is 2583 shots.
Samplers from #res4lyf works great.
#flux#lora#vhsevcore
Well, I can't even believe everything can be okay
The test was run at 512px because there wasn't enough memory for 1024, but the priority was just to get it running
#fluxdepth#fluxdepthlora#vibecoding
I am a pure "Chinese room" while I'm "vibe-coding"—a physical operator of copy-paste, moving from one window to another. I have no understanding of the internal workings or the consequences. I hope that willpower and brute force will pay off.
For two days I've been struggling to adapt a script to train a depth LoRA for Flux depth. I'm trying to port the mask channel logic from Sebastian-Zok's repo and integrate it into ostris's AI toolkit while accounting for how he handles FK the control image channel So far no luck
@MrForExample any chance to add fuction to preview 3d mesh on custom image input bacgkround and output for combined image output for futhermore processing? ^___^