Training question

[faint relic]

Hi I am just wondering how many steps and epochs i would need to set for 158 images, also what would be the best learning rate to set I'm using the Lora Trainer by Hollowstrawberry

[vernal zephyr]

There are a lot of disagreements about the best settings and everyone has their own opinions and preferences. This rentry has some simple fool proof settings that will help get you started: https://rentry.org/59xed3#the-easy-settings. It also had a ton of background information on various training methodologies.
DAdaptation is the rich kid optimizer that sacrifices vram for simplicity. If you're using hollowstrawberry's colab it should be fine.

Personally I'd recommend using either novelAI or anylora as your base model, it will make the lora more flexible when using it with different checkpoints

[faint relic]

Thank you, at the starting setting section it says 35-100 images for character and 100-10000 for style should i get rid of 58 images to make it 100 exactly?

[vernal zephyr]

no, more is better

though quality is better than quantity as well

generally you just want to hit the target step count. 1000 images could work, though the model may not learn the character as well if it introduces too much variety.

[faint relic]

I'll start to play around the settings and see what works thank you!

[vernal zephyr]

👍

If you want to do more tweaking there are a lot of other lora guides with different settings, you can find them on reddit and civitai. Dadaptation is a fairly limited optimizer with very little room to tweak since the learning rate is automatic.

[faint relic]

Is there a place to download the novel ai model, id like to use it in my webui

[vernal zephyr]

https://huggingface.co/acheong08/NovelAI/blob/main/stableckpt/animefull-final-pruned/model.ckpt 🤫

it's not really a taboo anymore though

[worn plume]

atlassian has stated that nai diffusion 1 is no longer copyrighted/they will not sue you

[faint relic]

little update and a few questions, I decided to do local training since it has more settings to play around with, and so far I got some good results. I trained it on 512,512 but i would like to get some more detail do you think it would be worth it to train 768,768 i tried it but it said it would take 13hrs I have a rtx 3060ti 8gb

here are two images that turned out well

also for some reason the eye colour doesn't apply unless I put it in the prompt, (it is in the activation tag)

[vernal zephyr]

640x640 is a good middle ground with bucketing since it will give you a 512x768 bucket which is good for inference. I'd also recommend disabling Nvidia's vram offloading since it tends to swap early which is probably where the 13 hour training time came from.
There's a guide for that here: https://nvidia.custhelp.com/app/answers/detail/a_id/5490/~/system-memory-fallback-for-stable-diffusion
The option was added in a fairly new driver so you may need to update.

[faint relic]

One last thing for some reason the lora likes to face the character's body away from the camera would that be a data set problem?

some examples

[vernal zephyr]

yeah that's probably the dataset, could be the model you're using for inference as well

like if it's already a bit biased to that pose training on more of it will make it happen much more often

[faint relic]

I'm using anything v3 i know its a bit outdated but its the only thing that seemed to work😓

should i go with v5?

[vernal zephyr]

I'd recommend https://discord.com/channels/930499730843250783/1096508811667325019 as a replacement for the anything models, there are a ton of others though. Generally more stylized models don't work with loras as well