Using open-source models to solve a core creation pain point in live-action comic drama production: face homogeneity in character design, making it possible to create vivid characters with genuinely distinct faces.
General workflow:
1. Use Anima Preview for text-to-image generation to quickly produce anime-style character concepts with diverse styles. The model is fast and supports a wide range of visual styles, helping establish a distinctive character silhouette from the start.
2. Flux2Klein offers very fast inference. Combined with open-source LoRA models such as anytorealcharacters/xhs realistic style and other diverse face datasets, it can precisely convert anime characters into realistic human textures.
Flux2Klein conversion is efficient and smooth, while the LoRA ecosystem is well-developed. It can freely call differentiated facial features, expressions, and temperament traits without complicated debugging. This open-source workflow breaks the template-like limitations of AI-generated faces at the foundation, giving each comic drama character a distinct identity and greatly improving recognizability and production efficiency.
Example:
Prompt preparation:
Close-up of the upper body. A beautiful East Asian woman wearing exquisite traditional Hanfu, with peach pink and champagne as the main colors. The chest and skirt are decorated with delicate floral embroidery, showing the unique sheen and texture of silk. Her hairstyle is a high bun, paired with jade hairpins and floral ornaments. She stands elegantly in front of a clean white background.
Close up of the upper body, a beautiful Eastern woman dressed in exquisite traditional Hanfu, with peach pink and champagne colors as the main colors. The chest and skirt are adorned with delicate floral embroidery, presenting the unique luster and texture of silk. The hairstyle is a high bun, paired with jade hairpins and floral headpieces. Elegant posture, standing in front of a pure white background.
Anima Preview generates a character sketch and fixes the rough visual identity. What we actually need is the face data from the LoRA, so other details do not matter much. On a 5090 with 32GB VRAM, one generation takes about 9 seconds.
Flux2Klein anime-to-realistic-human conversion can use prompts such as young Asian girl realistic style, middle-aged Asian woman realistic style, and elderly Asian woman realistic style to complete transformations across different age groups.
You can also see that Klein anime-to-realistic conversion takes only about 3 seconds per generation, making it very suitable for creating face models. After generating enough suitable face models, you can expand the image into four-view character sheets for comic drama creation.
If you do not mind generating several times to stabilize character consistency, Klein can also be used to generate four-view character sheets:
