Tiled NVIDIA PiD in ComfyUI

追加: 2026-06-01

[00:00:01]Hello guys, I know you listen about Nvidia Ped. This is fast and higher resolution latent decoding with pixel diffusion.

[00:00:15]This model can upscale latent images up to 4K but it has some problem. It can upscale only square images and uh but resolution is about uh 5,000 11 by thou 5,11 or larger we can fix it by using tile method of decoding. Latent images. All you need you need install comfy UI first and uh download custom nis. Custom nis can be downloaded uh on GitHub.

[00:01:21]put it in your custom node directory on config UI and download from example workflow folder. Um uh tile p example Jon else you need download uh some models from confi or we need download p flux one uh 5,11 to 2,00 48 BF is not it's safe tensor Um after you need to go to your config UI uh then go download the JSON file.

[00:02:16]Put it in your config UI.

[00:02:19]Uh here uh you need to insert your prompt insert your resolution res any generate small image by set image turbo.

[00:02:44]You need also download that image turbo model.

[00:02:48]Then it split latent image into tails and uh go to uh p diff diffusion it generate multiple image and connect it in one. Let's try press run generation start.

[00:03:17]We just look the result of Y standard decoding images and the PD decoded images.

[00:03:34]Generation finished it. Tail upscale started.

[00:03:40]Let's wait.

[00:03:43]Generation and upscaling finished it.

[00:03:45]Let's view result.

[00:03:49]Uh let's zoom in.

[00:03:59]You can see that image not changing but have a lot of details added.

[00:04:09]Where is that is eyes? We can detail it. See that reflecting it.

[00:04:20]Uh, it's very good. It's very nice.

[00:04:28]Let's try another prompt.

[00:04:36]Copy your prompt and paste into text ring multi-line generation has been finished. Let's see.

[00:04:56]Let's see on face first.

[00:05:02]Oh, that's cool.

[00:05:07]Wow. Oh, hair is perfect.

[00:05:13]Tourist is also perfect.

[00:05:23]Wow. We can see threads on clothes and we can see grain on rocks.

[00:05:39]That's not We can see small artifacts, but it's good.

[00:05:59]Let's try another prompt.

[00:06:16]Let's see how much generation took. It took,92 seconds.

[00:06:26]That's enough.

[00:06:29]Let's see the result. This is a woman portrait.

[00:06:34]Let's zoom in.

[00:06:43]I think that is good idealizations.

[00:06:50]Hair visualizations also perfect.

[00:06:56]Skin texture is perfect.

[00:07:00]Um flowers on the rest is also good.

[00:07:10]Aberations is added.

[00:07:21]Let's try last and we uh see how generations is processed.

[00:07:38]First of first step is text encoding. Text encoding use clip text encoder prompt. Negative prompt is not add any effect on this because um CVG is set to one.

[00:08:04]It's just keeping car sampler use SD sampler and simple scheduleuler we generate with constant noise seed we that set seed field generations took eight steps it uh Uh finish it soon.

[00:08:40]We preview small image low resolution. The resolution of full image will be uh will be seen here.

[00:08:57]It's finished.

[00:08:59]Generate preview of small image.

[00:09:02]Uh next is text encoding prompt for bit model tail for this model is constant size.

[00:09:20]This is for twice larger that source image source image is 5,1 output image will be 2,98.

[00:09:45]We produce four steps for each tail.

[00:09:50]We have about 11 tails.

[00:09:57]uh tails after decoding will be smoothly connected between it has overlap small overlaps and adjust brightness each tail to hit the um artifacts.

[00:10:33]Tails go from left to right and from up to down of the image on the image.

[00:10:44]Each image will be simply decoded by V.

[00:10:49]V is uh not loaded. It's just simply real RGB pixels model generates RGB format and just generate simple image as bitm or other formats.

[00:11:19]Uh last time I depoding we wait for it images will be mer by this and we compare with the image comparator. Resolution of output image will be about 35 megapixels.

[00:12:04]Decoding has finished.

[00:12:08]Let's zoom in.

[00:12:11]We can see that image more or contrast.

[00:12:17]Wow, that's perfect.

[00:12:21]Dilization is very good.

[00:12:26]We can see a lot of detail of on the image in unfocused areas detail is also added noise edit I think That's very good.

[00:13:01]Thanks for watching.

[00:13:04]Subscribe to the channel. Download custom notice links provided in description.

[00:13:14]If you like video, comment it. Ask for your questions.

[00:13:22]Bye-bye.