- This topic has 33 replies, 9 voices, and was last updated 1 year, 2 months ago by
.
-
Posts
-
I wanted to try my hand at something that SD does best...faces!
So I took some great Daz Vendors/Characters to see what SD would make of it...Enjoy!
Tyler HD G9 (from bluejuante) (AI)
Tyler HD G9 (from bluejuante) (Daz)
Kei HD G9 (from Mousso) (AI)
Kei HD G9 (from Mousso) (Daz)
Ravenna HD G9 (from Mousso) (AI)
Ravenna HD G9 (from Mousso) (Daz)
Victoria 9 HD (AI) - This is G9's flagship..right? 🙂
Victoria 9 HD (Daz)
These are all very nice and portraits is something that SD excels at. It's when you want to render two different characters doing something other than looking at viewer is where SD really struggles. Openpose and couple Loras help a lot but you'll have to do dozens and dozens of attempts before you get something acceptable. At some point I think that issue will be made easier.
A challenging art comp idea would be to use SD to have two different looking characters interacting with each other, SFW of course 😌.Very nice @legolas18 what checkpoint are you using?
@littleevie I am using
RunDiffusion XL
https://civitai.com/models/120964/rundiffusion-xl@frank22 Unfortunately you cannot prompt to have different characters (yet, but seeing how fast AI is moving, maybe by the end of 2024 we will have that).
I guess I can try to img2img something like that. I already scheduled to do something like it 🙂
In theory the BREAK command gives more control over how a string of prompts are interpretated so could help with generating 2 different characters.
For instance, if you want want a woman with red hair, a blue dress, and green eyes. You might prompt- "A red haired woman with a blue dress, and green eyes, looking at the camera".
A lot of the times when you generate that prompt, one or more of the three colors will exchange in the scene, eg. red dress or her eyes blue or maybe green. To prevent that from happening you can use the BREAK command like this- "A red-haired woman BREAK with a Blue dress BREAK and green eyes BREAK, looking at the camera"
I'm unsure if BREAK can be used to the extent of distinguishing 2 people but it's kind of thing that's needed to control prompts.
With Roop in SD you can actually designate a number to a character during the face swap. Being able to do this suggests it's quite possible to implement that method into SD to restrict prompts to a specific character in a multi-character scene.@frank22 I guess you could also use the AND logic expression 😉
Thankfully, SDXL is much better at following color prompts, although it does mess things up a bit.
I guess the ideal way would be to do img2img of a render or of a photobash (or a mix of both) 🙂
Yes for now but the easiest way would to be able to (like Roop) tag a character with an identifying number or letter (left to right) so that prompts can be confined to that chosen character. If that's already possible using Roop and ReActor it must be possible to do the same within SD with an extension, eventually 😕.
Ravenna and Victoria 9 daz render better than the AI. In my opinion. This shitty artificial stupid put that brow shape and texture for almost every girl.
Yes in some cases you're right but that can is down to the Lora (or lack of one) that was used to generate the images in img2img. If just "girl" or "woman" is the only prompt used, or none, they will all have the same basic face/eyebrows, etc.
Next time be more constructive than "shitty artificial stupid" - No soup for you... come back one year 😣.@legolas18
Just playing with the multiple character idea and very basic prompts.
Very diverse results and frustrating at present.
Took quite a while to get these 2 images.
Not really suited for easy replication of what's-on-your-mind right now, but interesting to see what happens.
As per the discussion - hopefully developments will add realism and easier work-flow to the creativity process.
@Frank22 Yes, maybe If you a prompt exprert and If you using embeddings, LORA's you will get more variations.
@adelasia Are you know the meaning of friendship? Did you read the books from Lucius Annaeus Seneca or Marcus Tullius Cicero about friendship? No. You have no friends. You have pals.
As someone who is training daily models since a year and a half (started in DiscoDiffusion) and is working for private companies, etc. in terms of AI trainings I will recommend you to use ControlNets instead of img2img. Img2img even at low denoise actually changes too much the thing, ControlNet nope, it maintains a good ratio of preserving original source vs quality-changes.
If you want more clear images or details on certain areas Adetailer extension works very well too (along as the typical face improvement ones). I'm more on the side of painting/concept art rather than just render 3D front faces but ControlNets will help you out.
Also remember that even in the inferencing (generations) you will need to adjust the CFG scale to the "style/aesthetic" you want. Example, realism in SDXL base probably 2-3 is good to go. In Juggernaut XL I think 2-3 works too. If you want more colorful images you will need to get a wider range using CFG scale 4 onwards to 7-8.
If not send me message or contact me via my username in CivitAI or Reddit. LDWorks David
Hello all,
My today PoV
Architector was a "little" rude as said Adelasia.
But..
After looking at a lot of AI portrait, maybe we could see, not a limit, but a tendancy.
AI make a average of their database weighted by the prompt used.
That's how they works in the end.
Results are never the same, but have so common traits.
That 's why they are so recognizable.
The AI touch ! 😉
For sure, it is the beginning of this aera.
And we could expect more refinnement soon.
But that also imply more complex prompts.
I'm wonderingg myself ...
How I've to spend my time to have the result I want ?
Modelization or prompting ?Not sure of the answer.
Dan.
- You must be logged in to reply to this topic.
