Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

[rayanarman]

The line causing the issue:

ComfyUI-HunyuanVideoWrapper/hyvideo/diffusion/pipelines/pipeline_hunyuan_video.py

Line 317 in 5291557

latents = latents * (1 - latent_timestep / 1000) + latent_timestep / 1000 * noise

I added the following before that line to fix it:

latents = latents.to(device)
latent_timestep = latent_timestep.to(device)
noise = noise.to(device)

[kijai]

With what workflow exactly? I couldn't replicate this with I2V, T2V or V2V workflow examples with current version at least.

[rayanarman]

I'm doing a 2-step generation using a low-res generation first, then upscaling the latent and providing that latentt to another HunyuanVideoSampler node on the input for samples, with denoise set to 0.6.
If I set denoise to 1.0 on the second HunyuanVideoSampler then error would not occur but below 1.0 was the error.

[eas125]

I have the same thing. Works if the second sampler has denoise set to 1, but results in 2 different vids with pixelation.

[eas125]

To clarify, this is t2v, not i2v.

[kijai]

I have the same thing. Works if the second sampler has denoise set to 1, but results in 2 different vids with pixelation.

It should work with the latest version of the nodes.

[eas125]

Sorry, still getting it. Deleting and cloned new repository didn't help.

[kijai]

Sorry, still getting it. Deleting and cloned new repository didn't help.

Can you try again now, it's working for me at least with latest push.

[eas125]

Looks like you got it. Tested with and without riflex just in case. Thank you.
I am getting an oom error with 720 and higher resolutions when using riflex now, but I'm not sure if that has anything to do with the sampler. 4090 card.