Good find! I’m having some success with creating characters such as in the OP. Somewhat less success if trying to describe a human character as I find you need to be much more descriptive with their characteristics (which, in turn, leaves you a bit less wiggle room to describe activity in the rest of the prompt)
Thank you so much Paul. I’ll give that a go this evening. I’ll let you know how I get on.
1 Like
Thanks Elmstedt! I was wondering how best to leverage the Customer Instructions. I’ll give that a go.
2 Likes
One note I’ll make. This is great for having a pretty consistent style but it doesn’t really work for perfectly consistent character design.
1 Like
Concur. I was able to get close, but there’s obvious differences if you look at them all together. The tech is coming along, though. I bet DALLE4 will have it… we can hope. Until then, I’m gonna keep tinkering with prompts.
1 Like
Here is what I am imagining we will see relatively soon-ish (2 to 4 years tops).
- Upload a reference image of a person or a sketch (or just a good text description of a character you want to create)
- Ask for a character model sheet to be generated of that character including various poses, facial expressions, and details of the character’s design, from multiple angles to give a comprehensive understanding of the character’s appearance and personality.
- Put that character model sheet (and others) into context for the model to reference when creating new images of that character.
I’ve not been paying too much attention to what the Stable Diffusion folks have going on these days, but even a year ago they were doing amazing things with consistent character generation, so I am sure this is probably already a thing or very close to a thing.
nexa
23
Actually there is a way to do this, but I think you’ll need to cross with midjourney. Here is youtube video on the process:
2 Likes
nexa
24
DALLE3 definitely does have its limitations in terms of graphical generations and photo manipulation, but midjourney is good for what you’re looking to do. So you can sketch or draft your characters on DALLE3 then upload image on midjourney to create a consistent character you’re looking for.
Yeah, I’m sure there are plenty of ways to do it.
I remember people fine-tuning LoRAs for individual, reproducible characters in Stable Diffusion upwards of a year ago.
Unfortunately, any commercial entity with deep pockets that wants to provide access to a generative image AI needs to fear the two threats of copyright holders and deep fakes.
So, no matter what the models are capable of there will need to be guardrails in place that will have the effect of reducing their utility.
nexa
26
You can request the seeds of the image and use that to recreate similar images with different variables.
for example you DALLE3 creates a superhero then you request the seeds from DALLE3 then you’re able to take the seed to recreate similar images or characters:
{
“prompts”: [
“Illustration set against the backdrop of Santa Monica Beach under a starry sky. The beautiful girl, with a clear face and no glasses, looks on with a mix of concern and awe. Beside her, the superhero is in a powerful stance, fiercely tearing away his tuxedo to reveal his superhero cape and uniform underneath. The scene captures the moment of transformation, with the fabric of the tuxedo rippling in the air and the emblem on his superhero uniform shining brightly.”
],
“seeds”: [4202394079],
“size”: “1792x1024”
}
but if you create a new chat and use the same seed it gets pretty wonky… so i recommend completing the character or goal with the consistency you’re looking for in one chat. I’m sure there is a way to use that same seed one month later to update the character, but you may need to tinker with it from then
FREG
27
Hey!
Just saw this long post in X trying to do this exact thing. Haven’t tried myself yet but try it out and see if it works.
Thread by @chaseleantj (Chase Lean)
1 Like
Looks like it might work if I ask Dalle to produce a grid of cards
1 Like
Foo-Bar
29
A grid prompt always works.
Create a “four panel comic about a fluffy bear” and the bear will be the same inside those four panels.
You can iterate on that image by using the gen_id of the image and refer to it by its referenced_image_ids parameter.
Great article here on Medium title : 99% character consistency with DALL-E 3
Can’t post a link here but I’m sure Google will find it for you.
Yup. I’m working on expanding the prompts now to see if I can stretch it to 24 scenes.
1 Like
I’ve found that sometimes DALL-E will produce these grids of images if you ask for two variations on an image.
For example, in this case, I asked DALL-E to “Give me two variations on that image.” I expected it to give me two images, instead, the output was a single image, but divided in two (below).
1 Like
A grid of images is really interesting, and also sparked an idea:
Ask it to create a grid with 16 panels and then upscale them with another AI (for now). Should get you 16 high-resolution images.
Bing DALLE3:
a grid that consists of 16 panels, each showing a teddy bear from different angles, in the style of Roald Dahl
Nero AI Upscaler:
Of course, not perfect since it leaves some artifacts, but now you have 16 images in the same style. The question is if all 16 descriptions will fit into one prompt…
Foo-Bar
35
With another upscale algorithm you can get even clearer results;
I always use the open source, stand-alone Upscayl, it runs even on a low-end tablet (no GPU power needed).
Upscale that upscale!
The image above is an upscale to 1024 x 1024, but you can even make it bigger with multiple passes;
(just a detail, I don’t want to blow up this server - the output was like 15.680 x 15.232 pixels)
Workflow
You lose some grain / details / accents, but when you prepare the image “to be upscaled” at forehand (make it a bit more graphical / slightly vector-like) then upscaling is not a big deal afterwards.
If you want to print it (magazine, not glossy - so 300dpi) the original image can be more than a square meter in size.
The preview in that image is it's left eye @25%.
Upscale that upscale!
In my opinion this leaves too much artifacts. It’s big, sure, but very AI-ish. Looks like the teddy bear is tearing apart.
1 Like
