"civitai" removed all my best examples because the source was real photos/videos. You can get better results from your own/other people's photos.
To UPSCALE the image use:
DENOISE: 0.015-0.15. (the (further the object/smaller the object) in the frame, the more it is subject to change). 0.015-0.02 to maintain a person's recognizability.
SAMPLER: euler for images with artifacts and poor detail. res_2s (RES4LYF nodes) for better images.
You can also connect 2 samplers in a row, first euler, then res_2s.
For i2i you can use any denoise power, you can even generate images by adding a small picture, which will allow you not to use the t2v_high_noise_14b model, which will save you up to 14GB of RAM/VRAM.
For v2v if the denoise is too high, the frame logic will be lost and you will get a "deforum".
Unfortunately I can't test it at a higher resolution because of 8GB of VRAM.
If you train Lora on the face, you will be able to perform a very high quality ROOP, even with objects that cover the face.
https://civarchive.com/models/1817671?modelVersionId=2057100 - t2v_low_noise_14b
Description
FAQ
Comments (7)
t2v_low_noise_14b model which one? or is there only one?
https://civitai.com/models/1817671?modelVersionId=2057100 ,one model is required and that's it, it is responsible for rendering. high noise is responsible for the logic of constructing frames as I understand it, and is not required for processing
viennar_ you rock!!! thank you
Wan2.2 is a miracle local video upscaler, and I'm shocked how few workflows here focus on this functionality. Many use the 5B model for speed, rather than half of the MoE 14B one.
The only downside is the 5 sec limit- when wan2.2 VACE is released, it should be possible to maybe triple that and keep temporal consistency.
wan2.2 is the best in my opinion, it upscales/retouches images. The workflow for video is additionally set because it just works.