Krea2-R lokr 克系美学
Back
Krea2-R lokr 克系美学
Krea2-R lokr 克系美学
Krea2-R lokr 克系美学
Krea2-R lokr 克系美学
Krea2-R lokr 克系美学
Krea2-R lokr 克系美学
Krea2-R lokr 克系美学
Krea2-R lokr 克系美学
Krea2-R lokr 克系美学

Krea2-R lokr 克系美学

60 Views1 Likes7 Favorites
2DFantasyStylization

Red gaga

Red gaga

2DFantasyStylization

Model Information

Active
Model Type:
LoRA
Basic Model:
krea2
Trigger Words:
克系美学,数字化作品,猩红面具,
Resource Name:
models/loras/Krea2-R lokr 克系美学.safetensors
MD5:
90eb313c68c8b7f5bfdb10dd8f293646

Besides trigger words, please keep prompts under 30 words! Short descriptions, anime style auto-invasion.

Recommended for use in this workflow. More efficient and saves points.
https://www.RunningHub.ai/zh-cn/post/2110636832210817025

When used alone, a weight of 1 is sufficient. Since it leans towards anime invasion, using some realistic words directly might result in a more live-action look. When adding the trigger word: Cthulhu aesthetics, you can append words like digital artwork, scarlet mask, etc., afterwards to enhance the anime style.

When combined with face models, a weight of 1 is generally sufficient. Please refer to the last sample image for optional weights.

LoKr (Low-Rank Kronecker Product): Uses Kronecker product for decomposition (ΔW = C ⊗ (B·A)) (e.g., splitting 512×512 into 16×16 and 32×32). The Kronecker product has multiplicative rank properties, meaning LoKr can achieve a higher equivalent rank than LoRA with controllable parameter counts by designing the dimensions of the factor matrices, thereby introducing block-level structured repetition patterns. This structure makes LoKr typically smaller in parameter count and file size than LoRA at the same rank.

Possesses extremely strong expressive power and higher parameter efficiency. Under specific configurations, the parameter structure of LoKr is even more compact than LoRA, capable of capturing richer visual details and cross-dimensional complex interactions.

Training stability is medium, more sensitive to data quality, and usually requires high-quality, diverse training sets. It is especially suitable for generation tasks requiring style features, structured layouts, and complex spatial relationships (such as style transfer), as well as scenarios where model size needs to be significantly reduced in resource-constrained environments.


Recently found a great AI image and video generation tool RunningHub! Supporting infinite canvas, Agent quick generation, and ComfyUI, you can freely combine AI applications, workflows, and generation tasks, with Seedance 2.5 starting as low as ¥0.21/sec. Open the link: https://www.RunningHub.ai/?inviteCode=rh-v1681 Register now to get 1000 RH Coins and generate quite a few images and videos for free!

Besides trigger words, please keep prompts under 30 words! Short descriptions, anime style auto-invasion.

Recommended for use in this workflow. More efficient and saves points.
https://www.RunningHub.ai/zh-cn/post/2110636832210817025

When used alone, a weight of 1 is sufficient. Since it leans towards anime invasion, using some realistic words directly might result in a more live-action look. When adding the trigger word: Cthulhu aesthetics, you can append words like digital artwork, scarlet mask, etc., afterwards to enhance the anime style.

When combined with face models, a weight of 1 is generally sufficient. Please refer to the last sample image for optional weights.

LoKr (Low-Rank Kronecker Product): Uses Kronecker product for decomposition (ΔW = C ⊗ (B·A)) (e.g., splitting 512×512 into 16×16 and 32×32). The Kronecker product has multiplicative rank properties, meaning LoKr can achieve a higher equivalent rank than LoRA with controllable parameter counts by designing the dimensions of the factor matrices, thereby introducing block-level structured repetition patterns. This structure makes LoKr typically smaller in parameter count and file size than LoRA at the same rank.

Possesses extremely strong expressive power and higher parameter efficiency. Under specific configurations, the parameter structure of LoKr is even more compact than LoRA, capable of capturing richer visual details and cross-dimensional complex interactions.

Training stability is medium, more sensitive to data quality, and usually requires high-quality, diverse training sets. It is especially suitable for generation tasks requiring style features, structured layouts, and complex spatial relationships (such as style transfer), as well as scenarios where model size needs to be significantly reduced in resource-constrained environments.


Recently found a great AI image and video generation tool RunningHub! Supporting infinite canvas, Agent quick generation, and ComfyUI, you can freely combine AI applications, workflows, and generation tasks, with Seedance 2.5 starting as low as ¥0.21/sec. Open the link: https://www.RunningHub.ai/?inviteCode=rh-v1681 Register now to get 1000 RH Coins and generate quite a few images and videos for free!