Qwen Image 2.1 Anime2Real — Alpha
A new Alpha version focused on achieving a truly realistic, photorealistic appearance from anime-style images, including more realistic skin texture.
Notes
No trigger word required.
No longer trained exclusively at 1:1 aspect ratio, improving compatibility with a wider range of anime-style images.
Compared with the previous Semi version, this model focuses on true photorealism, including more realistic skin.
Expression conversion has also been improved to a certain extent.
You can further fine-tune details such as eye color, eyebrow color, facial expression, and other features through prompts.
Recommended LoRA weight: 0.8
Alpha version.
Try it online：https://www.runninghub.ai/zh-cn/post/2104937924847812610/?inviteCode=368034ba/?inviteCode=368034ba
Qwen Image 2.1 Anime2Real — Alpha
全新的 Alpha 版本，这次不再是 Semi 半写实路线，而是专注于将动漫风格图像转换为真正的写实、照片级效果，包括更加真实的皮肤质感。
注意事项
无需触发词。
不再局限于 1:1 比例训练，对不同构图和动漫风格图片的兼容性有所提升。
相比之前的 Semi 版本，本版本更加注重真正的写实效果，尤其提升了皮肤的真实质感。
对表情转换也有一定程度的提升。
可以通过提示词进一步微调眼睛颜色、眉毛颜色、表情以及其他细节。
推荐 LoRA 权重：0.8
Alpha版本
Qwen Image 2.1 Anime2Real-Semi — Beta
A beta test LoRA for converting anime-style characters into a more realistic / photorealistic appearance.
Notes
Trained at 1:1 aspect ratio. Results tend to be better when the original image has realistic human head-to-body proportions.
NSFW is supported.
You can fine-tune the result by adding details such as eye color, eyebrow color, facial expression, etc. in your prompt.
Recommended LoRA weight: 0.9
Trigger WordConvert to realistic style
Beta version — experimental / testing use.
Workflow: Qwen2.1_Anime2Real-Semi_Beta
Try it online: https://www.runninghub.ai/zh-cn/post/2102935762902974465/?inviteCode=368034ba
Qwen Image 2.1 动漫转半写实 — 测试版本
用于将动漫风格角色转换为更加真实、写实的外观。
注意事项
模型按照 1:1 比例训练。如果原图接近真实人物的头身比例，通常可以获得更好的效果。
支持 NSFW。
可以在提示词中加入眼睛颜色、眉毛颜色、表情等细节，对生成结果进行微调。
推荐 LoRA 权重：0.9
触发词Convert to realistic style
Beta 测试版本，主要用于实验和测试。
Description
Qwen Image 2.1 Anime2Real — Alpha
A new Alpha version focused on achieving a truly realistic, photorealistic appearance from anime-style images, including more realistic skin texture.
Notes
No trigger word required.
No longer trained exclusively at 1:1 aspect ratio, improving compatibility with a wider range of anime-style images.
Compared with the previous Semi version, this model focuses on true photorealism, including more realistic skin.
Expression conversion has also been improved to a certain extent.
You can further fine-tune details such as eye color, eyebrow color, facial expression, and other features through prompts.
Recommended LoRA weight: 0.8
Alpha version.
Try it online：https://www.runninghub.ai/zh-cn/post/2104937924847812610/?inviteCode=368034ba
Qwen Image 2.1 Anime2Real — Alpha
全新的 Alpha 版本，这次不再是 Semi 半写实路线，而是专注于将动漫风格图像转换为真正的写实、照片级效果，包括更加真实的皮肤质感。
注意事项
无需触发词。
对不同构图和动漫风格图片的兼容性有所提升。
相比之前的 Semi 版本，本版本更加注重真正的写实效果，尤其提升了皮肤的真实质感。
对表情转换也有一定程度的提升。
可以通过提示词进一步微调眼睛颜色、眉毛颜色、表情以及其他细节。
推荐 LoRA 权重：0.8
FAQ
Comments (14)
Am I missing something - I always thought that Alpha comes before Beta but I'm getting the impression that your Alpha version is later than your Beta??
I think that is because Beta is "semireal" and Alpha is from a complete other branch bein "Real"
Haha, I know 😂
I just feel like “Alpha” sounds more like the original / mother version. So I went with it. 😆
@Masmalkar Exactly! 😂
Beta = Semi-real branch.
Alpha = Real branch.
And somehow, “Alpha” just felt like the mother/core version to me. 😆
Well, it works - I tried it.
It is common knowledge that the Alpha is stronger than the Beta, and the Prototype is stronger than everything else. 😈
@larflarf Just like the T-virus. 😈
food for thought: if I feed the original image through a tagger, filter the 2d,anime, illustration, censorship tags from the list, and pass it ahead to the prompt, would it yield better results?
The core idea of this LoRA is the conversion itself. I wanted to leave some freedom for the prompt, so you can tweak things like eye color, eyebrow color, facial expression, skin texture, etc.
There’s not really a need to completely rewrite the image with an i2i workflow — that would make things a bit too complicated. 😂
@arpalest I dont mean as a complete rewrite, but have the tags exactly as helpers for things like those details like the eye color, expression, etc which would definitely be pulled from the tagger, instead of the user having to edit the prompt for every image they want to process
@kaizerwp Ahh, got you now! 😂
Yes, that actually makes a lot of sense, especially for batch processing. Using a tagger to automatically extract things like eye color, expression, etc. and feed those tags into the prompt could save a lot of manual work.
In most of my example images, I actually didn’t use any additional descriptions, and the LoRA handled them pretty well on its own. Occasionally, I would add something specific, like “hair covering one eye,” “one eye closed,” or a particular facial expression.
So your approach could be a really useful way to automate those small adjustments. 👍
Is this super Asian biased or can it generate other races, like Caucasian, well? Tried so many of these that only do Asian I figured I would just ask at this point before wasting time struggling.
It largely depends on the art style. I also have Caucasian characters in my posts, you can check them out. Most Japanese anime art styles naturally lean toward Asian facial features. You can try American cartoon styles instead.
脸的重复度太高了，只能降低权重来避免都生成同一张脸