Wan 2.2 img2img workflow for the GPU poor

(the images in the showcase are comparisons of the techniques, left: original, right: Wan2.2)

for more info about the masking technique, read the article about focal masking.

(image found on reddit)

Wan 2.2 is different from its predecessor and virtually every known open source model. There are 2 different models used to generate videos while relaying each other back and forth until the process is complete. The models are referred to as high-noise and low-noise. The high noise controls the composition while the low noise does the details. At 14 billion parameters, these models are more capable and are equally more resource intensive.

Quantized models are a good alternative to regular ones which aren't always optimized for lower-end gpus. They're often the only suitable models for older gpus. Wan 2.2 is definitely a good candidate to be quantized. Because the models are trained on videos rather than images, they have a better interpretation of temporal space (how object positions relate to one another) and the detail is consistent and coherent. As a result, video models are often more superior than the image ones.

An img2img workflow using such models will conform the details of an image to a frame in a video. After all a video is just a sequence of images. In this case you will have to use the text low-noise model as that model is responsible for creating the latent space to generate the image. The other models will not work as well as that one. As with any img2img workflow: the lower the denoise level, the lesser the change would be in the output. So adjust the level accordingly to your desire. Also resolution is crucial as it affects the detail. However the generation time increases with resolution to the point where Comfy would run out of memory.

For nsfw images the results are predictable: it mangles genitalia.

I had only tested the workflow with nsfw images on men only. It often turns penises into a foot or some other vague appendage. I have doubts that it can preserve the details as vaginal closeup as well. The only remedy that I know is inpainting it.