To avoid any potential legal risks, please do not buzz me.
请不要给我buzz。
If your initial image has poor results due to disproportionate head-to-body ratios, you can increase the strength of the high-noise LoRA—for example, to 1.2 or even 1.5.
如果你的初始图片头部和身材比例失调而导致效果不好的话,可以增加高噪声部分的lora强度,例如1.2,甚至1.5
This is an anime-to-real (A2R) LoRA. I haven’t seen this type of LoRA on Civitai, so I decided to create one myself—this way, I don’t have to keep making separate cosplay outfit LoRAs, although I still think the results aren’t quite as good as dedicated outfit LoRAs.
The low-noise part of this LoRA likely needs more training, but I don’t have time for that right now. The high-noise and low-noise components were trained on different datasets. Initially, I used the A2R dataset from XiaoZhi (Jason), which was based on Qwen. I converted real photos into videos using Wan 2.2 i2v, then used Adobe Premiere Pro (PR) to composite the anime images together.
However, since Wan 2.2 isn’t an editing model, when trained on diverse source materials, the resulting model struggled to maintain consistent hair and clothing colors. To address this, I gathered a new, curated dataset and trained a fresh model. Unfortunately, the skin texture turned out quite poor, so I ended up combining the previous low-noise weights with the current high-noise ones—and the result is actually pretty decent.
All training data excludes any uncensored content. When generating images with special poses or actions, you can combine this LoRA with other LoRAs for better results.You can combine this with other i2v or t2v LoRAs available on Civitai to achieve specific poses or actions.
这是一个动漫转真人的lora,我没有在civital看到这种lora,所以我自己做了一个,这样我就不必再去制作cosplay的服装lora了,虽然我认为效果仍然不如服装lora。
这个lora的低噪声部分可能需要更多的训练,不过我暂时没空,高低噪是使用不同的数据集训练的,我最开始是使用小志jason大佬的qwen的a2r的数据集训练的,我将真实图片用wan2.2i2v做成视频,然后用pr将动漫图片拼在一起。但是由于wan2.2不是编辑模型,在素材种类多的情况下,训练出的模型无法控制发色和服装颜色的一致性,所以我自己找了一批素材炼了新的模型,但是皮肤质感很糟糕,所以我用之前的低噪搭配现在的高噪,效果还不错。
训练素材不包含未经审查的内容,有特殊动作时,使用的时候可以搭配其他的lora。
你可以搭配civital上其他的i2v或者t2v lora来完成一些特殊的动作
Description
FAQ
Comments (24)
What's the prompt scheme? Or do I just need to insert a "keyword" and everything will work?
Am I doing something wrong? I input an anime clip, used the "动漫转真人" trigger in the positive prompt, but the output is still anime, no change.
I'm not sure about the exact reason, but are you using the original model?
@qiaoy123 Thank you for the reply. I am using Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf
@lamentcounterbalance this is an i2v lora,you should use it on i2v model
@qiaoy123 Strangely I still get the same result even with Wan2.2 I2V 🤷♂️ Would you mind posting a workflow? I would appreciate it.
@lamentcounterbalance I'm not using my own workflow, and the one I'm using is very basic—almost identical to the official workflow—so I don't think it's particularly meaningful. However, you can share your workflow with me.
@lamentcounterbalance This LoRA has another issue: if the character's body proportions in the anime image are severely distorted, it will simply follow those incorrect proportions. So that might be the reason.
@qiaoy123 Yes, I see. I think that was the issue. I tried a different anime and it worked!
Is this the same lora as https://civitai.com/models/1710040/realistic-transformation?modelVersionId=1939608
The linked one is Wan2.1
If I had known this existed, I wouldn't have trained my own LoRA.
I think mine has better consistency, but it's not as versatile as this one.
@qiaoy123 i just spent basically an entire day tuning a workflow to make that wan 2.1 model work without totally changing the image, with some success but a shit load of messing around. this seems to work much better. thanks.
@qiaoy123 wan2.1 Lora can work with wan2.2 but they often degrade image quality. It's good you made your own 2.2 one
if you find this doesn't change the image as much as you'd like (still looks anime) then try turning the shift up and the cfg.
Thank you for your sharing.
V1 and V2 work excellently, doing the job as described. Well done.
Why is it zipped?
Can you go from real to anime as well?
Actually, there's already a LoRA like this based on the Wan2.2 version available on CivitAI—you can try searching for it.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.