This works ok! even understand gags well enough.
training notes: dataset 260 @ 1024, long qwenvl3 tagging, 64/64, 0.0002
Description
FAQ
Comments (10)
I had not much luck doing loras for zimage, so far. Can you give me a hint about the VL3-Prompt worklfow you use for captioning?
qwen vl3? i mean it's nothing special: Describe the image in extreme detail for ai image dataset... Focus on... Start with... Mention...
i noticed an improvement when going from very simple or even danbooru captions to very wordy qwenvl3 captions. but maybe im imagining that or it was something else
@Joschek Sorry, I meant, whether you can give me a link to the actual workflow? The ones I tried gave results, but z-image didn't really "get" what to do with the result.
@ragnaroook not sure what workflow you mean. i caption running llama.cpp. i vibecoded my own small captioner if you want to have a look.
@Joschek Thank you very much, I'll give it a try
Your parameters should also matter! I see that you are using a different LR than typical. I also see that the number of steps seems quite high. LoRA metadata say epoch 4 and steps 19200!! (so for a set of 260 that would be around 19 repeats? )
Are these correct?
Well, although I consider ZIT inferior to Flux and Qwen (especially qwen) it has some advantages (not only speed) and who know what the next versions will bring
@zkrd im Training Here online and step count seems to be calculated differently to previous models. So I'm not actually sure if that's correct. I think this was epoch 9. LR is the one setting I've I've seen noticable correlation to dataset size + quality before. So far would say inconclusively that 0,0002 seems to be a better starting point irrespective of dataset size
@zkrd to me prompt adherence is on another level compared to flux1 (qwen and flux to are too slow for my taste). There are some aestical things which are not perfect but that's what style lora are for.
@Joschek True! Prompt adherence is quite good (better than Flux, comparable to qwen, maybe better sometimes). Have I understood correctly that you used civitai trainer?
@zkrd yeah, training locally bricks my laptop for too long