Mirrored from https://huggingface.co/wikeeyang/Magic-Wan-Image-V2/tree/main

Magic-Wan-Image V2.0

This is an experimental model and a blended version of the Wan2.2-T2V-14B text-to-video model. The goal is to allow Wan 2.2 model enthusiasts to generate various images using the Wan2.2 T2V model as easily and conveniently as using Flux.

The Wan 2.2 model excels at generating photorealistic images while accommodating various styles. However, because it evolved from a video model, its generalization capability in image generation is slightly weaker. This model strikes the best possible balance between photorealism and stylistic flexibility while capturing as much detail as possible, virtually reaching a level of creativity and expressiveness comparable to the Flux.1-Dev model.

The blending method for this model involves layer-wise mixing of the High-Noise and Low-Noise parts of the Wan2.2-T2V-14B model using different weight ratios, followed by light fine-tuning. As an experimental model, it may still have some limitations. You are welcome to try it out and share your feedback so that improvements can be made in future versions.

Features realistic skin textures and ultimate high definition, making it ideal for generating portraits and real-world photographic images. Rich in detail, it supports direct upscale via TTP up to 8 megapixels. This model is a clean base model; users can mix and match various NSFW / SFW LoRA models according to their personal preference.

Usage: Please refer to the example workflow image: V2-example-workflow-image.png

• model shift: 1.0 - 8.0, feel free to experiment;
• model cfg: 1.0 - 4.0, feel free to experiment;
• inference steps: 20 - 50, feel free to experiment;
• sampler / scheduler: deis/simple or euler/beta, or any combination—feel free to experiment.

Also on: ModelScope, Civitai

The model is under Apache 2.0.