27.01.2026:
No matter how much easier that would have made things – there is no single best AI image model out there right now. There will never be.
So, the "Qwen superiority" shift has been revoked in favor of the flexibility of multi-modality. Although Flux is no longer my preferred model, I am now trying out Z-Image for my characters. I love its candid vibe and ultra realistic look overall! I think this model fits my work well. So, if I decide to make a "shift" again, it'll be more like this:
Flux LoRas of OCS – no longer upgraded because they've become unpredictable and unreliable
Qwen LoRas of OCS – passively supported
Z-Image LoRas of OCS – main.
21.01.2026:
I have decided to make a big shift. All of my original characters with their own LoRa models (Kira, Tori, Tera-byte, Iris, Evelyn, etc.) will now primarily be supported on Qwen instead, as this model has shown far superior world understanding compared to Flux and some others. I was bewildered at how badly it could mess up basic anatomy, or how heavily it generalizes certain types of characters. This especially showed when I was trying to train Kira's 4.1 Flux LoRa. She looked "whitewashed" and had abnormally disfigured, almost animal-like hands and facial features. Additionally, I acknowledge that Qwen is much harder to run on consumer GPUs, which I can use in my favor like this:
My OCS will require extra hardware to run, which makes sense, because I don't want to let just about anyone use them.
However, I will continue to create style LoRas, public character LoRas for the community on Flux,so that those could be used by people locally. I think this is a pretty good compromise.
20.01.2026:
4.1 release successful. Nothing too extraordinary, but I'm glad I could add some more data.
18.01.2026:
4.1 release FAILED. Looking for a resolution soon.
17.01.2026:
4.0 out now, upgraded to 96 2k resolution images!
Available with an E1 and E2 option. Not sure which epoch is best yet, but some hand distortion on Epoch 2 (although it resembles her face the most) concerned me, so I made the Epoch 1 available as well. I have a feeling that it will be the best.
It's really so weird though, because epoch 1 was only at 960 steps... You really can't predict when a model is going to be good and when it will start to deteriorate... It's all just trial and error.
04.01.2025:
3.0 retrained in 2k resolution for Flux. Qwen model will no longer be upgraded. Z-Image-Turbo LoRa experiment coming soon.
3.0 for Qwen release log:
Same hyper-realistic dataset, even more opportunities. Compatible with both regular Qwen and Qwen Edit! Qwen appears to be the best model for realistic virtual influencers.
3.0 release log:
More realistic and detailed, much more consistent, upscaled greatly on the face, and made to look a bit older as she has canonically aged. Best to be used with Krea's Flux!
Original description:
Kira is my OC who's gone through a lot of redesigning... This is her final form, hopefully not as confusing as any previous one... (Many people have been asking me if she's CGI or maybe a doll ever since she started to look more realistic, lol)