Thanks, There’s hardly any info about Draw Things. So everything you post is really a valuable resource. How many images do you use for the training? Or what do you say is the minimum or maximum that makes sense. Also, would you suggest training of 1.5 Lora for people with only 16 GB of ram or use SDXL with 512 x 512? Thanks for the great content.
I think you can do 512x512 XL on 16GB, but it will be slow because the memory will sometimes need to swap to the hdd. Shut down all other apps, especially web and Discord. number of pics is maybe between 5 and 25 (mine was 27 here), more pics will increase RAM so use your best images.
I think so, just reload the LoRA by name and the settings should fill-in as before. Then you can change the photos and set the training steps higher to resume... But also consider it might be easier to start a new training session and create multiple LoRA - rather than over-training a model you like with extra garbage.
I get best results from head/face shots (include eyebrows to top of shoulders) different angles, expressions, and lighting/environment. My pics are cropped square 2048x2048, rendered from 3D software.
Hi, thank you for this tutorial and all your content. Could i use vertical and horizontal different resolutions?, for example, 384 x 576 px (for portrait images)
I have only ever trained with square images... I have heard that people do use different ratio images while training, but I think all the pics need to be same ratio.... I think the trick is to continue training by adding more steps with all new images - but again, I have not tried it.
Thank you for the really helpful tutorial. I tried to train my own LORA with 12 photos and managed to get the captions, but every time I start the training the program crashes. My Mac has a i5 Intel processor with 8GB RAM and a 8GB Radeon card. Do you think training LORAs will be possible if I increase my RAM or do I need a M1 or a M2 processor from Apple? What is your advice? Thank you in advance.
I was hoping someone with more hardware knowledge would answer..., My guess is the older architecture will not be able to handle the GPU requirements, without the new fancy Apple Silicone 'unified' memory. I still recommend DreamLook.ai for training, it is fast and a lot cheaper than a new Macbook...
That happened to me too. The app crashed on my iPhone and iPad as soon as I started training. I then switched to my MacBook M2 and the training went through