Closed xueqing0622 closed 2 months ago
Sorry about that. Would you be able to provide the error message? or is it just not behaving as expected?
if you could provide the workflow file that might help me troubleshoot as well :)
The hiresfix node has lanczos In rescale method.. it is intended to be used with an upscale model, and then rescale that. if you dont want to use a model beforehand, the sampler upscale should do the trick.
there are no error message, just use lanczos method up total mess, and other upscale method also no good in ttn pipeKsampler
here is the workflow file: ttn-sample-withUpsacle.json
I find some lanczos upscale nodes here, but lantent upscale still not good effect, use image to upscale and vaeEncoder better than lantent upscale. https://githubfast.com/Extraltodeus/CustomComfyUINodes/blob/main/LatentByRatio_not_mine.py
pipeKSampler
enlarges latent
instead of image
. It is completely different from the general high-definition repair method. Use hiresfixScale
to get the results you want.
I think we can add a upscale by model
function to pipeKSampler
to optimize the experience
Of course, I look forward to your updates, and I'm going to be very critical.
haha, just a joke!
Plz update for new pipeKSampler. maybe can add a image input, to upscale then vaeEncode to latent in the pipeKSampler, I found a lot of upscale for latent give us bad result.
It seems that you didn't understand. What I said adding upscale by model
to pipeKSampler
is equivalent to integrating hiresfixScale
into pipeKSampler
. The upscal effect is the same as webui, and lanczos can be used normally.
In the final analysis, the incorrect image generation is not caused by pipeKSampler
, but because you used the wrong method.
The upscale methods within pipeksampler use the default comfy upscale code.
As mijuku said, using hiresfix is the preferred method for a "true" hires upscale..
I'll consider adding model upscale into the ksampler, but it is already starting to get a little crowded.. we'll see
Model upscale has been added to pipeKSampler_v2 in cfaecd01e7f3547b38c7e44f417655471a6b181a
I like to use the lanczos algorithm, because it smoothes out the zoomed pixels without smearing. but when use lanczos in ttn pipeKsampler, it show error result,
And how about to add lanczos to upsacle model in hiresFixScale
just like SDwebui, I just need lanczos only, other model zoomed pixels will smearing