Magic-Wan-T2IV V3.0
***网络问题，请不要下载本模型文件，烦请移步 HF 或 ModelScope 下载***
Due to network issue, Pls Download from : Huggingface.co, Modelscope.cn
模型在 v2.0 的基础上，有以下几方面的提升 (Has the following improvements:)：
更逼真的实物景致与人像肌肤，更好的提示词遵循能力。(Realistic scenes and skin, better prompt compliance)
更好的美学构图与人脸优化，对 NSFW 支持能力更好。(Aesthetic composition and facial optimization, better support for NSFW)
文生图与文生视频统一模型，支持较短长度的文生视频能力。(Unified for T2I and T2V, support short-length T2V capabilities)
支持英文及少量中文字体的渲染。(English and a limited of Chinese fonts Rendering)
本模型依旧为纯净底模，可用于微调或 LoRA 训练，也可叠加各种 Turbo / Distill / NSFW / SFW 的 LoRA 使用。
Pure base model, suitable for fine-tuning or LoRA training, and can also be combined with various Turbo / Distill / SFW / NSFW LoRAs.
感谢：@亮亮rayne，@GuangyuanSD，@kanttouchthis，@QuantFunc 在本模型调制优化过程中给与技术、构图、美学、量化等方面的指导和支持。
感谢：@T8star-Aix，@AI-KSK，@黑鹤001，@大麦茶看AI，@joeMultimedia 等一众大佬对本模型的技术评测、使用介绍、社区推荐及进一步微调提升等方面的建议和支持。
基础工作流请参见(Base Workflow)：V3-T2Image-Workflow.json, V3-T2Video-Workflow.json
==============================================================================
Magic-Wan-Image V2.0
逼真肌肤、极致高清，适合人像与实景照片类图像的生成。细节丰富，支持 TTP 直放到 8M 像素。本模型为纯净底模，用户可根据个人喜好组合添加各种 NSFW / SFW 的 LoRA 模型。
Highly realistic skin texture and ultra-high resolution, suitable for generating Portrait and Real-world scenes. rich details and supports direct upscale to 8M pixels using TTP technology. This is pure base model, you can add all kind of NSFW / SFW LoRAs according to you like.
基础工作流请参见：https://civarchive.com/images/109925926
Base workflow pls refer to: https://civarchive.com/images/109925926
Also on: [Huggingface.co](https://huggingface.co/wikeeyang/Magic-Wan-Image-V2), [Modelscope](https://www.modelscope.cn/models/wikeeyang/Magic-Wan-Image-V2)
==============================================================================
Magic-Wan-Image V1.0
本模型是一个实验模型，是 Wan2.2-T2V-14B 文生视频模型的混调版本，目的是能让广大 Wan 2.2 模型的爱好者，能像使用 Flux 一样，简单方便的使用 Wan2.2 T2V 模型来生成各种图片。Wan 2.2 模型擅长写实类图像的生成，同时兼顾多种风格，由于是视频模型演化而来，模型在生图能力的泛化性方面稍微弱一点。本模型最大的兼顾了模型的写实能力和风格变化，同时尽可能的体现更多的细节，基本上达到与 Flux.1-Dev 模型相当的创作力和表现力。模型的混调方法是将 Wan2.2-T2V-14B 模型的 High-Noise 和 Low-Noise 两部分分层进行不同权重比例的混合，再简单微调而成。目前是一个实验模型，可能还存在一些不足之处，欢迎大家试用并反馈信息，以便在未来版本改进。
This model is an experimental model. A mixed and finetuned version of the Wan2.2-T2V-14B text-to-video model, Let many enthusiasts of the Wan 2.2 model to easily use the Wan2.2 T2V model to generate various images, similar to use the Flux model. The Wan 2.2 model excels at generating realistic images while also accommodating various styles. However, since it evolved from a video model, its generative capabilities for raw images are slightly weaker. This model balances the realistic capabilities and style variations while striving to include more details, essentially achieving creativity and expressiveness comparable to the Flux.1-Dev model. The mixing method used for this model involves layering the High-Noise and Low-Noise parts of the Wan2.2-T2V-14B model and blending them with different weight ratios, followed by simple fine-tuning. Currently, it is an experimental model that may still have some shortcomings, and we welcome everyone to try it out and provide feedback for improvements in future versions.
More quantized version Pls refer to: https://huggingface.co/befox/Magic-Wan-Image-v1.0-GGUF , Tks for bewolf jason's work.
使用方法：
请参见例图工作流，https://civarchive.com/images/98412305
or Please refer to Workflow: https://civarchive.com/models/1931055
model shift: 1.0 - 8.0，您可以自己尝试；
model cfg: 1.0 -4.0，您可以自己尝试；
inference steps: 20 -50， 您可以自己尝试；
sample / schduler: deis/simple 或 euler/beta，或任意组合，您可以自己尝试。
The model under Apache 2.0.
Description
逼真肌肤、极致高清，适合人像与实景照片类图像的生成。
Highly realistic skin texture and ultra-high resolution, suitable for generating Portrait and Real-world scenes.
FAQ
Comments (49)
Have anyone done a workflow for inpainting using this model?
I found a solution that works great.
Share it, Bro, Thanks...👍
@wikeeyang https://pastebin.com/5MmsMxcE
@wikeeyang https://civitai.com/posts/24469290
Did a fast upload gen, just to upload it. I played around with it yesterday and it worked great. It takes some experimenting to get what you want. Like strengh on ksampler and such. Typical AI gen stuff haha
@noyboy Thanks a lot for share your experiment...👍
@wikeeyang No worries! :)
Its not perfect, but for my use it works good. i can use my loras and put face on. And then edit all the rest in krita :)
@wikeeyang do you think its possible to make a magic wan, but with wan 22 animate. And get the power of controlnet (depth and pose)? 🤔
@noyboy I don't konw, I think for refine inpainting should be ok, for mask area or character keep inpainting may need some lora or controlnet to support.
@wikeeyang hahah I see now that my question came out like a request, but it was more a shower thought. :D
@noyboy thanks, but as I know, Wan2.2 and Wan 2.2 Animate have different model structure. that can't mix together simply...
模型很厉害，我尝试自己把示例工作流改成了二次采样放大工作流，出2K分辨率的图，细节棒的很，就是速度稍微慢了点。cfg1.0，step10也可以出完成度较高的图片了，可以节约点时间。
分享您的经验，谢谢！另外，这个模型可以直接用 TTP 放大到 6 -8 百万像素，且细节不丢失。👍🤝
Thank you. Possible to have the full fp16 or bf16 weights?
Yes, but it's still in the experimental stage and doesn't feel mature enough yet, so it won't be released for now.
@wikeeyang ok, thanks.
im quiet new to this i have a few questions.
1. is it possible to train character using this magic model or should we train character using wan 2.2 base model ?
2. is it possible to train and copy art style with this model ?
3. is it possible to generate new image using an image refference character for example change the pose of the refference image ?
thankyou in advance !
Should be ok, you can try it use wan2.2 trainer, and low noise lora of wan2.2 model may get better compatibility.
does it wok on forge ?
太完美了，使用体验完全超出了预期，请问大佬能做个I2V版本的生图模型吗？因为I2V有生图能力，但我即便把像素开到100w，33帧，切镜后参考出的视频还是比较糊，能不能做到像文生图一样或者比原本画质更加清晰锐利呢？如果能做到的话，wan将会对任何图像编辑模型做到真正的降维打击
哈哈，感谢您对本模型的肯定，没试过 I2V 的生图，有空时倒是可以试一试。。。🤣🤣🤣
This person stole your model https://huggingface.co/aquif-ai/aquif-Image-14B
Thanks a lot! Despise him! People like him sully the spirit of open source.
@wikeeyang It seems he has deleted his org, lmao. I have screenshots of the threads from HuggingFace anyway, haha
@wikeeyang It's one of them https://files.catbox.moe/bfy666.png read it
@qek Thanks a lot! Bro, I will save the screenshot image as a memorial, I hope him can fully recognize his mistake and learn from it. Whether it's HF, Civitai, Modelscope, Github, or any other platform, countless technical researchers and creators are sharing their technologies and models. We should respect the hard work of researchers and creators, be grateful for their willingness to share everything, and maintain an attitude of communication and learning to study and share together. This is the only way technology (including AI) can better serve all of humanity. This is the spirit of open-source technology, and we must embrace all open-source work with honesty, a willingness to learn, and gratitude. Finally, I wish we together to remember: The technology can't tolerate any deception, deception will not make ourselves happier, nor will it make the world a better place.
@wikeeyang Everything I was able to screenshot https://files.catbox.moe/ytyltm.7z
Have a nice day
@qek Thank you very much for you integrity, kindness and help for me.
@wikeeyang aquiffoo the thief removed Magic Wan from his collection named 'My top picks'
@qek Thanks a lot, let's forget it. Let it go... 🤣🤣🤣
What other CLIP models work with this version?
It should working ok, but i don't test it detaily.
Do WAN LoRAs (high noise / low noise or any) work with this model?
Both work
This is beautiful, I can use my own trained LORAs for WAN, WAN's clip & VAE with this to create sharper crisper Image.. Love it.
Thanks for your like this model. 👍🤝
In my testing I've found this to be a direct drop in replacement for the stock wan 2.2 refiner (highnoise). And in this regard it is superior. It will create exactly the same image just more sharper and detailed.
Thanks for your detail testing for the model.
这个就是wan2.2的low模型，可以生图也可以生视频，也就是说任何low模型都可以这样用，哈哈哈This is the wan2.2 low model, which can generate images and videos, meaning any low model can be used this way, hahaha.
对不起，兄弟，这不是 Wan2.2-Low-Noise 模型，说话要有事实依据。/ Sorry, Bro, This not the Wan2.2-Low-Noise part, Pls show with facts as the basis.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.