An experimental LoRA for Qwen-Image-2.1 that converts manga and anime character images into realistic, photo-like human depictions.
Trained with paired manga and realistic images. It aims to retain the source composition, pose, clothing, and scene while changing the visual appearance.
Tested settings:
• LoRA strength: 1.0
• Sampling steps: 25
• CFG: 1.0
• Sampler / scheduler: Euler / simple
• TextEncodeQwenImage21 resolution: 1536 — approximately a 2.36-megapixel area budget, rather than a maximum edge length.
• Image-aware prompt enhancement was used during testing.
Results depend on the source image, workflow, prompt, and seed. Facial expressions, closed eyes, winks, mouth shape, iris color, and small accessories may vary. Some results have smoother skin and a more stylized appearance.
This is a Beta release. Strength 1.0 is a tested reference setting, not a proven optimum. Feedback and comparisons with the base model are welcome.
Description
FAQ
Comments (3)
what is the prompt?
Hi! I recommend trying the ComfyUI workflow included with my example images. It contains the actual prompts and settings I used for those results.
In my experience, using the training captions directly as generation prompts doesn’t give the best results. Please use the example workflow as a starting point and adjust the prompt to suit the image you’re converting.
transform the image to realistic photograph