Download
1 variant available
1230 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
10 1 2 3 4 5 6 7 8 9
30 1 2 3 4 5 6 7 8 9
(21)
Aug 4, 2026
BETA
Anissa – Invincible | Krea2 Beta One LoRA, many styles: 2D show · Cosplay/Realism · Game · Comic · Chibi 3000 steps · 78 images · Strength 1.2–1.4 Turbo-friendly (8–12 steps) · Works with other LoRAs – share gens & feedback!
Show more

670 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
2410 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
2.3K0 1 2 3 4 5 6 7 8 9.0 1 2 3 4 5 6 7 8 9K
Anissa LoRA (Krea2 Beta)
Anissa from Invincible – Krea2 Beta LoRA Trained for 3000 steps on a carefully curated 78-image dataset.
This is an early but highly capable beta model designed with one clear priority: versatility.
Instead of locking Anissa into a single art style, this LoRA was trained to handle a wide range of visual languages while keeping her recognizable. You can generate her in:
Classic 2D animated / TV show style (this makes up the largest and most refined part of the dataset)
Cosplay / photorealistic looks (trained on multiple different cosplayers so the model produces fictional/AI people only — it will never recreate a real person)
Video game aesthetics
Comic book / sequential art style
Chibi and stylized cartoon versions
The same character, completely different vibes, all from one LoRA.
Because other characters from the Invincible universe were deliberately included in the training data, this model also works extremely well when combined with other character LoRAs. You can place Anissa next to other heroes or villains and keep consistent style and anatomy across the whole scene.
Technical notes & recommended settings
Best strength range: 1.2 – 1.4 (still usable at 1.0 if you prefer softer influence)
Performs very well on Krea2 Turbo — high quality results in just 8–12 steps
Works across every common aspect ratio
Always include quality tags (masterpiece, best quality, highly detailed, etc.) for maximum clarity and polish
Current limitations (this is a beta)
The dataset is not yet fully or consistently captioned, so prompt engineering still matters.
Results can vary depending on how strongly you push certain styles.
This is the first public version — further training and a much larger, higher-quality, fully captioned dataset are planned.


