A VAE for Wan2.1 and Krea2...
Inspired by JoeyGambino and his Z-Image + H3/LTX merge/grafting process. Using one of the highest benchmarked VAE currently available as the mathematical basis for modifications to the existing WAN VAE and it's 2x Upscale variant.
For the Krea2 Upscale version use this node pack to run it:
https://github.com/spacepxl/ComfyUI-VAE-Utils
minimax version 1 is just a pixel shift modification to the existing VAE. Version 2 adds additional detail and also more accurately matches start-frame input images in detail, colour and lighting.
I will work on adding additional details to it if possible.
For minimax use this nodepacks vae decode:
https://github.com/TripleHeadedMonkey/ComfyUI-MiniMaxH3_LatentUpscaler
If the VAE still doesn't work for you (size mismatch error) then download the optional zipfile included and add that to "\ComfyUI\comfy\ldm\minimax\" replace the file vae.py" in this folder.
Description
No additional details added with this version, just upscales the output 2x without needing additional models. Perfect for a resampling workflow.
FAQ
Comments (3)
note sure if I'm doing something wrong, but to me the 2x upscale for Krea2 just makes everything look washed out and ridiculously oversharpened.
I'm already using the required nodes, but are there any other adjustments necessary to make it work or is just supposed to look like that?
This version is currently better for artwork than realism. I'll work on a better version in the near future. But the method is experimental and requires a lot of testing by hand.
Also it is a model-based thing. if the model was fine-tuned to look sharp using the original VAE, it will look oversharp on the new VAE. Whereas a model that was smoother, will now look normal.