(Please forgive the mess, using Civitai to host these LoRAs with different training settings to compare.)
Enhances the lighting to be less studio and more interesting. A work-in-progress.
Trying some LoRA captioning techniques.
Dataset 1 (48 images)
- v 0.1a - short captions
- v 0.1b - long captions
- v 0.1c - short + long captions (48 images + 48 same images)
(Dataset 2 (176 images)
- v 0.2 super moody
0.1a-c have fewer images in the set and are more moody darker images with film grain.
0.2 super moody trained on 4.5x as many images with more variety, with 2 copies of each image, one with a shorter caption and one with a longer one.
Try the trigger word moodlit for more control.
Description
FAQ
Comments (2)
Which was most accurate to the way you envisioned your prompt in your testing?
Duplicating the images with two sets of captions (short + longer) seemed to work well. Will definitely keep that as part of my process moving forward. In terms of total training, a lot of guides on Krea 2 say to aim for fewer more-perfect images but I still prefer in this case the results from the larger set. This is also a style LoRA and a lot of guides are made by people doing characters so that may be part of it.
In terms of the total steps I used, I had Claude Fable research this pretty extensively and there just wasn't very good information out there so we made some guesses. In all cases we seemed to have overtrained and going back a few epochs and reducing LoRA strength seems to produce better results.
I'm happy with how the current versions are working, but gonna run another batch soon with some changes and see what happens. Once I have enough data to feel like I've learned something, I'll write an article about it as well. Krea 2 is such a cool base checkpoint. Really excited to see where we can take it.