- Dec 21, 2020
- 1,575
- 3,776
Happy new year! Yea, my money on 2024 being (2020+2022)^2 tooHappy new year!
View attachment 3221884
Lets hope this year sucks less and is less crazy.. Not gonna hold my breath though
Right after the Davos meeting ends (January?) We'll have the new current thing right around February, down to the week: ~week 3, same as the last two current things.
- Jan 24, 2020
- 1,401
- 3,803
- Dec 21, 2020
- 1,575
- 3,776
FML, that's a sweet lineup. One for each day of the week and two for Sunday please. Monday rather, two for Monday please.
- Dec 21, 2020
- 1,575
- 3,776
E-xc-el-le-nte! Thank you for this.
I am re-doing the dataset replacing "strokes" with "gradients". The going is slow and prolly will take a week or so. Planning for this look below to be the default for painted clothes - that should be v2. If v2 looks good all in all, then v3 should have detail such as pockets, zippers, seems, and so on.
- Jan 24, 2020
- 1,401
- 3,803
Remember to caption the type of clothes she is wearing so we can hopefully have some control over this and the color as well. Anything that is not captioned is then just a part of the general image generation and is outside of our control. I know it's tedious to edit each manually but it is well worth it in the end. Many loras are bad exactly because people take shortcuts like sloppy or even incorrect captions.
Last edited:
- Dec 21, 2020
- 1,575
- 3,776
1000%. Yes, this is the plan for v3. The v1 was to see if the thing is workable altogether. It proved the concept and showed me that I am not hot on the strokes, but I rather prefer the gradient.
Now, I want to settle on the actual look in v2 - but the dataset won't yet have enough details to caption. Reason is I do each piece manually using vector drawing. This takes ~15-30 minutes a piece (between the actual shape and the aesthetics) where the v1 was 1 minute at best cause I just sprayed the canvas. Probably the bult of output is gonna be shorts and hotpants, just as v1.
But v3 I definitely want to have blouses and jackets, so those will come with details and much more detailed captions.
- Dec 21, 2020
- 1,575
- 3,776
What's your opinion on throwing highly detailed but partial pieces into the dataset? Like this (let's pretend this is high quality render of painted hotpants only, although naturally the image I am presenting is a mere crop of a fullbody shot):
- Jan 24, 2020
- 1,401
- 3,803
It depends on the goal of the lora and the types of images you want to generate with it. Will you or do you want to generate images that looks like this and with this composition? The better the quality of the source images the better the lora and the resulting generated images. I would not include this if it was up to me, you have to decide for yourself though. Perhaps it will give better result than I anticipate.
- Jan 24, 2020
- 1,401
- 3,803
If you like the composition you can of course include it. Just be very clear in the caption. Don't include too many images like this because it could make it a dominate feature of the lora and a majority of the generated images has a cropped head. This is very frustration when you wish to have the face included. So only include 1-2 images like this and with very clear captions about the composition.
Preview of my upcoming Lora's for SDXL and 1.5
Completely new data set of 112 images all 1024x1024 for XL
XL Training Preview Image
Completely new data set of 112 images all 1024x1024 for XL
XL Training Preview Image
You don't have permission to view the spoiler content.
Log in or register now.
