"DALLE GIRL SCOUT AMBASSADOR" is a imaginary character born as experiment. She is completely artificial made. After I discovered a trick to have a consistency with DALLE (the image generator inside ChatGPT) I generated lots of images of this ScoutGirl, just to prove the concetp. After this, I realized the character will be lost when the chat will be closed and, at that point It would a pity. So I asked to user "Iselestia" to train a LORA for stable diffusion with my dataset of selected images... He was so kind to accept...
VERSION 1.0
Dataset made by me.
LORA training made by Iselestia
This version 1.0 is well trained and can do a very good scout dress and also a nice girl.
Unfortunately the train is not 100% succesfully and the face doesn't resemble to original charater. I hope this crucial point will be improved in future versions...
Trigger word: there is not. The dress is automatical generated but you can also change the clothes, specifiing some differentet prompt.
Compatible with all checkpoints I tryied... the original dataset from DALLE was REALISTIC, so it will have more resemblance with PHOTOGRAPIC or SEMIREALISTIC Checkpoints.
If you want a good looking tutu' skirt, I strongly suggest to add a tutu' LORA in the prompt, like this one: https://civarchive.com/models/19235/tutuballet-outfit
If you like a good model, I suggest to have a look at Iselestia profile:
Link to Iselestia models: https://civarchive.com/user/Iselestia/models
Description
VERSION 1.0
Dataset made by me.
LORA training made by Iselestia
This version 1.0 is well trained and can do a very good scout dress and also a nice girl.
Unfortunately the train is not 100% succesfully and the face doesn't resemble to original charater. I hope this crucial point will be improved in future versions...
Trigger word: there is not. The dress is automatical generated but you can also change the clothes, specifiing some differentet prompt.
Compatible with all checkpoints I tryied... the original dataset from DALLE was REALISTIC, so it will have more resemblance with PHOTOGRAPIC or SEMIREALISTIC Checkpoints.
If you want a good looking tutu' skirt, adding a tutu' LORA in the prompt is suggested.
If you like a good model, I suggest to have a look at Iselestia profile:
Link to Iselestia models: https://civitai.com/user/Iselestia/models
FAQ
Comments (5)
What's the DALL-E 'trick'? (If it's not too valuable a secret)
Simple: every image you make with DALLE has a gen_id and a seed, (ask for them and ChatGPT will tell to you).
-gen_id is just a unique code to identify every single image in the chat
-seed is a random number used to generate different images every time.
The resulting image is due to prompt and seed, so if you keep the prompt and you manage to fix also the seed you will get exactly the same image.
Since DALLE is very precise, if you fix the seed and change the prompt just in the action done, you will get a different image but in the same style and with exactly the same character.
But there is problem: OPEN AI locked the feature to ask a seed value to DALLE for "security reasons" (read: paranoic censorship).
Here comes the trick (but have a bit of patience, now is late and must go to sleep, I will continue tonorrow)
Hello again. So, the workflow is to ask a prompt to ChatGPT, descriving the character with precision (age, hair, eyes etc...), and then let it generate random seed images, until you get a character you like, At this point, ask for gen_id and seed used in previos images. Then, since you cant ask directly to use a seed, just ask ChatGPT to add in the prompt for DALLE the part: "Reference to gen_id= 'XXXXXXXXXXXX' ", where the X is obviously the one you like, (copy and paste it). Make attention to use " and ' exactly as I wrote here, otherwise it can confuse the number. The funny fact is DALLE will be inducted to use the same seed of the image you referenced and you will check this success because it will generate only one image, infact if prompt and seed are the same, there cant be "variations". Continue asking to keep the reference but changing a little the action of the prompt: you will see the character is exactly the same, same face, same style... variations can happen only in the parts you forget to fix in the prompt. For example, with my scout girl I forgot to tell the eyes color, but there is no doubt I got 140 images of the same girl walking, climbing, runnig, sitting, coocking...
@settima_ai Thank you
@davy272 Thank to you.
(here, in a post were I put some original images, I put the link for the ZIP archive with original dataset from DALLE. It's all done with the "trick" I described to you. You can download and check by yourself how works substantially well )
https://civitai.com/images/6184301 (look in the text on the right)
By the way, please note there are 3 drawbacks:
- locking the seed, it will generate only 1 image at time, not two.
- if you delete the chat, the charater is lost forever, because there is not a way to use the seed in another chat. And the "gen_id" has a meaning only in that precise chat.
-For the same reason, you cant give your charaters to friends. It work in your chat only.