This is the Qwen-Edit-2511 version for anime-to-human conversion, offering more realistic results and higher generalization compared to version 2509. It's not limited to anime-to-human conversion and was trained using only 28 datasets.
You can click the link on the right to find the workflow
Recommended prompt:
Convert to realistic style or 变成真实风格Strength: 0.8~1.0
Currently, very few tested examples have failed conversion, and the generalization is decent. If you encounter any issues, feel free to leave a comment.
Description
FAQ
Comments (25)
will this work with 2512?
it just work for edit 2511
2512 isn't an editing model (yet?)
@Dead_Internet_Theory 2512 is a text-to-image model.
@shadowsii I know, I mean they may or may not release an updated -Edit variant
@Dead_Internet_Theory Never
@shadowsii If I could please make a suggestion or shed some light on this subject, please stop calling everything 2D as anime. it's 2026 and I hope people should know what "anime" really is by now. Not everything 2D or Illustrations can be blatantly named as anime. I don't care if this comes off as rude but this has to stop, I'm so bloody tired of people calling every damn thing as anime.
what?
Without any intention of offense, as I don't really understand these well, how should I refer to these types?
@GaryXAIGC these look like illustrations or 2D stuff in general. Anime/Manga is only referred to toons made in Japan, somehow over time people have completely misunderstood the actual meaning of anime and end up calling everything 2D as anime, which isn't the case. When you speak about it they get offended
workflow link not working btw
My oversight—it's now accessible again.
Found time to test it and it works great!
One question though: Is there prompting to keep the ethnicity of the input image like european/western looks?
The girls all have very strong Asian facial features, and images with a mixed diversity seem to end up in a group of Japanese girls, maybe if you decide to make a new version someday, add more diverse ethnicity into the dataset, hair color seems to be picked up very well.
Anyway, cheers for this lora which takes away some steps/headaches to come to a similar result!
In addition to specific trigger words, you may also describe the racial characteristics you wish to generate for the task. Realistic Lora types incorporating Western features can also be added, and these will all take effect.
@shadowsii Thanks, I tried that, but with multiple characters that does not seem to work (no idea how qwen 'counts' the characters (top/bottom, left/right, top left/bottom right, no clue as each gen it seems to be different).
@TheLastRemain I have tested that adding Lora is effective.
@TheLastRemain with multiple character, first type the main prompt convert to realistic then in bracket (explain what the image is about but mentioning clothes color or some identification, it works flawlessly.) like the girl in yellow dress is American / french,
I tried several different ways to transform non-realistic images (2D/2.5D/whatever) to realistic style and this LoRA with Qwen Image Edit 2511 is just amazing. All in all this is the best way to get realistic images of fantasy characters. Qwen Image Edit 2511 does not seem to be censored at all, it just lacks knowledge of certain "NSFW parts". But this can be easily fixed by combining this LoRA with a NSFW Qwen Version like Qwn-Image-Edit-abliterated v1.5 (for some reason, it works way better with v1.5 than v1.6).
The only downside of this LoRA is that the resulting realistic person often lacks facial expressions (which then has to be prompted for) and almost always looks directly at the viewer, even if the original 2D character doesn't.
If the resulting image is slightly shifted to the original, I recommend using the "ReferenceLatent" node in addition to the usual "TextEncodeQwenImageEditPlus" node for better adherence to the original image. This causes the generation to slow down but in most cases it is worth it.
PS: Sorry for filling up the LoRA's gallery like this. ^^"
Thank you for your feedback. Since I didn't prepare much data to train this LORA, as the current results are sufficient for my needs, I plan to add more datasets to Flux2-Klein for retraining. This model is simply outstanding.
Amazing Lora. I really like it! Thank you!
works great, but it completely ignores faces. like eye colors, expressions, features like eyepatches etc.
I created a workflow that compared the lora effects of multiple anime-to-real human conversions, including this lora（anime2real-2511）.You can click the link to give it a try.
Thanks to the author for providing such a powerful tool for our community!
With Qwen 2511 generates truly outstanding images. Even black-and-white comic images are rendered in color. Amazing! By the way, could you please specify the license for this model?
will it work with Rapid AIO ?
I tested it and it does. But you need to reduce the strength otherwise you might get artifacts. For me, depending on the source image, I use from 0.10 to 0.50.