Upscaling In Comfy UI Done Easy! 

Hey man, i loaded your json.. i have the issue that only 1 frame gets restored in the final output .. oddly, frame 3, the rest of the workflow seems to work - it puts the blurry mouth on the og footage - with just that 1 frame (frame 3) being restored by the model (either model, same effect).

That's an awesome tip thank you! I love using drag and drop and clipboards for being more efficient

you're a fkn legend sir

Thank you

That's awesome thank you so much for the comment!

Thanks for your feedback glad it helped :)

Thank you so much

Thank you so much

Awesome! when I first switched from Automatic 1111 my first task was to figure out upscaling in the way I required. Now I have been playing with all the different upscale models :)

Thank you for watching. Hope I was able to help a little.

No Problem! Hope it helped

It's such a solid model to be honest. Works pretty much for everything, no need for the refiner it's great!

Yep I almost have my new setup completed

So the way I do this is by setting the latent image width in the range of the trained image in the same aspect ratio. So I use a ratio plugin. Then from there I will upscale to the native resolution of the monitor or what ever use I may need. Now for ultra wide what I would do is go slightly above the ratio and crop down or smaller and crop in. Because the trained image is 1024x1024 you need be be careful going outside of that as it can cause tiling and rendering issues. Using known good ratios will solve this.

I would try both but likely after the upscale will yield better results. As for which are more friendly towards photos and illustration so far I find the 4x Ultra Sharp is great for illustrations. It can do well with images that are already a decent size but we are starting to see more and more AI based upscaling which uses more of an image generation mind set which can reduce the amount of noise. That said any fine details like text, or objects the AI can not detect turn out to look a little weird. Text is a great example, it turns our looking like an alien language if it's to small. This is because AI Upscalers just like image generation has not been trained on text and has no clue what it is, so in some cases traditional upscaling methods will still be required in photographic upscaling. I was a professional photographer for years and I remember always trying to push my raw images out as large as I could to see what would happen, trying to find new techniques to get the scale I wanted. In most cases for printing it was unnecessary as if your printing off a really large poster or doing a billboard for example. The viewing distance is further than say something like a 4x6 printed image in your hands. Now that said if you need a larger upscale than something like 4X it may be worth the time to generate in a larger size first and then up scale only that 4x. Another amazing trick is if it's an illustration like a vector image, once created bring it into Adobe Illustrator and use the Trace Function. Work with it to get the vector to what you want then you have a vector image you can scale to any size with out worry.

Can you please point me to a video for upscaling with adding details in comfy? THX!

If you only want to do straight upscaling, then ComfyUI is a slow and "expensive" i.e. GPU power and time... way to do it. Use an online or dedicated upscaler like Topaz or Upscayl and get faster and better results. The proper way to use upscale in Comfy or A1111 is to do Latent Upscaling, thereby adding more detail by utilizing latent space. Similar to HiRes Fix.

@@TheDocPixel Intresting you say that. I found and adapted a workflow that iterates adding detail as it scales, yes quite slow but what a great job it does. The larger I go the more detail it adds. Straight upscalers just don't work for me. They get bigger sure but something is always sacrafised.

@@uk3dcom - I agree... and also have adapted Detweiler's workflow. The results are stunning when also adding an upscale model.

​@@uk3dcomI am looking for a workflow that adds detail and not just upscale. What would you recommend?

I can start saving them for you guys and I’ll upload them to a git hub account I’ll set it up in the morning with my workflows :)

@@AIMadeEasier that's great looking forward to it

This is extremely likely. Upscaling requires a bit of horse power. You could try a lighter upscale model from that list and try and run it multiple times. Although not ideal. There is two reasons for a box to go red. #1 the node is not installed (up scaler is pre-installed) so we know it's not that. #2 the node had an error running (lack of v-ram, incorrectly configured ect...) Most stuff these days for Stable Diffusion they recommend 6gb vram or higher, I would not go below 8 personally. Depending if you have the budget or what not, it might be worth looking at a new card. A great example of a budget card is a 3070 used or even a 4060 as these cards are fairly inexpensive and work well with stable diffusion. If your on a laptop then once again depending on your budget might want to run in Google Collab or even using a cloud based solution. How ever with cloud based solutions, if you forget to shut it down can cost you the price of a 3070 used in a single months time. So just be aware that cloud based solutions can be extremely expensive. Example: Paperspace on their basic tier. The basic tier is $0.46 per hour so if you forget to shut down the server when not in use your monthly cost could be as high as $355. That said if you are using it lightly and you are diligent about shutting down the server than it may be more affordable. Just with the risk price of $355 and a used 3070 going for around $300 these days... You get where I am going.

thank you so much, I tried other Upscaler and it worked @@AIMadeEasier

@@sc0peAI yeah depending on the upscale will use more or less ram. If you can find a low vram upscaler that is almost lossless you can stack them. So 2x-2x until you hit your vram limit.

He has moved it to here tensor.art/models/642019060365894429

It depends what your end goal is. If you need a really large image; you may need to use something a little more powerful or in the cloud. So it really depends on your end goal. I am working on a video about running stable diffusion in the cloud and how to do it on a budget.

@@AIMadeEasier Not really large. Just doubling a 512 x 768 image. It takes me around 10 minutes per upscale, which is a bit longer than what I'd like. I don't know if there's a faster option.

No not really you would need to export all the frames of the video back to image then upscale each image and then recompile back into a video file. There are tools out there for video upscaling though that would be much easier.

Of course this tutorial is not for everyone. Yes using proper techniques you can do a lot with photoshop and photography tools. I use to be a commercial photographer and some of my images out of camera and on film were scanned in at 100mp or higher then upscaled for specific jobs. For the average person though, that is unnecessary. The point of the stack was to show that upscales can be stacked. So if you are just looking for an all in one solution this was one of the best methods at the time of recording. That said there are new upscaling techniques that have since provided better results and less compute intensive. But also trying to compare this to large format photographs is not the point of this video or in the scope I was going for in this video. I appreciate your input tough.

Sure it does this video was done with a 3070 8gb :)