Decode VAE producing yellow tint

[KadTableAttack]

ComfyUI VAE decode constantly produces a weird yellow tint regardless of which VAE, checkpoint model, CFG, sampler, scheduler, prompt I use.
I tried the same seed, steps, CFG, sampler, scheduler, model, VAE and prompt on WebUI Forge and it doesn't have this issue.
Does anyone know why this is happening? This is really annoying because I prefer ComfyUI.

I have:
NVIDIA GeForce RTX 5060 Ti
CUDA versio 13.0
Drivers version 580.76.05

[blepping]

That is a strange issue. Most likely it's related to your VAE, so two things you can try:

Try using the "FP16 fix" VAE from madebyollin (you'd put this in models/vae): https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

Get the TAESD SDXL decoder model here (this will also enable high quality previews if you turn them on): https://github.com/madebyollin/taesd/blob/main/taesdxl_decoder.pth - put that in your models/vae_approx directory. The built in VAE loader node has a taesdxl option you can use. Note: These are for fast (and lower quality) VAE decoding but you can use this to determine whether the problem is definitely at the VAE decode stage or not.

If you VAE decode with those and still have the yellow tint issue then it is extremely unlikely your issue is VAE-related.

[KadTableAttack]

Hi, thank you for the suggestions.
I tried both of these decoders and the result is the same. A yellow tint in the end image. It's related to ComfyUI since WebUI Forge doesn't have this issue with the same models/vae. So either there is some setting I need to additionally set in ComfyUI or it's actually a bug.

[blepping]

This is a very strange issue. The only possibly unusual thing in your setup from what you mentioned is using CUDA 13, I think most people are on 12. I'm using CUDA 12.9.

Since you now have TAESD for SDXL, can you try using that preview type? The preview type in your screenshot above is latent2RGB which is fast but very low quality. Using TAESD previews might help you see if a problem occurs during some point in sampling (and maybe somehow the tint issue isn't apparent when using latent2RGB previews).

We've seemingly eliminated VAE as an issue, but I struggle to imagine a case where you could see a TAESD preview that looks fine in the sampler and get the tint issue in your final issue when you're using TAESDXL for the previews and the VAE decode step. It seems like that would be impossible.

Are you starting ComfyUI with any commandline options? Using custom nodes? (They can have an effect even if you aren't using them in a specific workflow.)

edit: Actually, never mind what I said about CUDA 13, apparently I'm on CUDA 13 too. PyTorch doesn't have a version that specifically is for CUDA 13 though, they only publish versions for CUDA 12.6, 12.8 and 12.9 (I'm using the one for 12.9). Are you using the portable installation of ComfyUI?