Eider v3: Joining the (agentic) dark side
About joining the agentic dark side, getting burnt, and learning from it
It has been longer that I had expected since I trained Eider v2. v2 was ok but it had a problem - different environments (outdoors, close portrait, mid body portrait, indoors) provided significantly different results - simply because the training data contaminated with different faces for different scenarios. That is known, its fine.
The problem was how to get out of that reinforcing look, so future LoRA-s didn't entrench themselves more and more into that bias. And the generations for v3 were going well, relatively, but there were some looks and positionings that could just not be achieved. Mostly because I was being stubborn with only using SDXL. Unnecessarily.
The new version produces consistent faces, particularly indoors, with caption bases models such as ZImage or MOSS (Image below generated with MOSS) working very well at certain poses.
The first act - Giving up
After a painstakingly long process of curating a collection for v3, I eventually gave up. I was stuck in certain poses and certain biases could not be overcome. And, in a burst of frustration I deleted all the photos and said "that's it, I'm done with LoRAs"
Stupid, of course, but it was done. Because I deleted them with no recourse to recover them.
Then over time, having left the project aside for a while, two things happened. I noticed that ChatGPT's image generation within the chat was very sticky. Once it was stuck in a style, or shape, or face, It didn't like changing it. A gouache style digital artwork that had degraded to a different style was impossible to recover even if you used the exact same prompt. Which was a pain. Unless you were wanting it to be sticky. Like, with a character.
What if I fed ChatGPT enough "references" that it became obsessed with them? And, after trying it, this seemed to work quite well, surprisingly, but it also lost the plot eventually. Worth considering it, after all.
At the same time something else came up in the world, a new fashion: agentic models. I tried Claude Code for a while and did a few silly projects, such as simple browser game or a very simple music player.
But then I decided to try Codex, given I had paid for ChatGPT plus - or whatever it is the cheaper subscription. And I noticed that other than code, it could generate images autonomously.
So in an afternoon, I instructed it to generate a batch of 50 randomised captions or prompts (outdoors, indoors, close portrait, mid portrait...) And then generate batch images for those prompts.
Look! it's working!
And just like that, after leaving the computer working doe about an hour and a half or so, I had 50 reasonably good pictures or a very resembling character, each with an accompanying txt file with the caption.
And just like that, encouraged by human laziness, I had Eider v3 which was improved significantly since v2.
It is not all joy a happiness though. This has its own problems. Eider was generated from a single reference image. Because I was lazy, and I thought it wouldn't work. Being an external image (OpenAI's model has its own preferred faces) it took that face as the only true face, including perspective. So the vast majority were facing (her) right side.
This is visible in the results. I forced Codex to change faces and I did generate a batch. And for the SDXL LoRA I simply mirrored some images. Again, lazy.
The model also has its quirks. Outdoor photos have a very subtle painterly style - as if you could almost see brush strokes. This is fine, in a single photo from ChatGPT. They are actually impressive. But when you pack 50 of them together to reinforce a LoRA, you get the effect on steroids - and that can be noticed on outdoors generations.
That isn't terrible, it can be solved by editing the image - not a big deal if you already have the character right.
See the odd effect of the scenery? at first it looks fabulous, but when you generate 50 it looks weird. like paint strokes. ChatGPT likes it. And the LoRA puts it on steroids.
In any case, a lazy afternoon job generated a pretty decent character LoRA. And the main problems are - I think - solvable at source.
Having learned from all of this, I am already working on a second character, using other technics that I find interesting and which should definitely have the face direction problem well solved. The painterly strokes will be harder. But its good progress. And it opens many doors to different possibilities.