This is a great article, but could you host the videos somewhere with quality controls? It's hard to judge the difference between input and upscaled when the video seems to be playing back at a low resolution with significant compression artifacts.
I spent a moment trying to make sense of this infomercial...
Do I guess correctly that when you write "all in ComfyUI", you mean "all in ComfyCloud using paying cloud based services that have nothing to do with ComfyUI" ?
SeedVR2, FlashVSR, Wan2.2 Upscale, Fast GAN, and also traditional non-AI upscale workflows are all open-sourced and available locally! We also attached all the workflows at the end of the article. You can always download the workflow and run locally!
Sure, but you didnt attach downloadable workflows at the end of the article : you gave links to workflows on ComfyCloud. Whenever I click on the links, I get a message asking me to choose a paying plan. That does not sound very "open source" to me, even if the upsacle models are effectively downloadable elsewhere, if you manage to get the information from other sources !
An by the way : I really do appreciate the great open source work the ComfyUI team is doing... I was just a bit disappointed to find that you were pushing proprietary solutions and did not give easy access to the WFs.
We didn't link the workflows directly since some of them do require custom nodes and it might be confusing for people who don't have them locally already. We thought that it will be helpful if users can take a look at what the complete workflows look like and then they can decide how to use them.
Yes all of them are available on comfy cloud. The links at the end of the article will take you direct to the workflows on cloud. You can also choose to download the workflows if you prefer to use some of them locally.
I find Topaz is poor for illustration, smooths off the character of any natural media style. I find a tiled upscale with appropriate loras will do a better job for very high res.
the key to tiled upscales is to be able to prompt and denoise individual tiles . Ultimate upscale doesn't allow this so you get hallucinations in tiles where nothing much needs to change, a sky for example. I've done up to 16,000 pixels across but you need to add more tiles so that each tile is the ideal size for generating, IE about one mega Pixel for Flux.
Thank you for the before/after comparisons!
Thank you!!!
thanks for this one! Super informative and I was pulling out my hair trying to get behind this.
This is a great article, but could you host the videos somewhere with quality controls? It's hard to judge the difference between input and upscaled when the video seems to be playing back at a low resolution with significant compression artifacts.
Good point. We can at least just share all the files. Will it be helpful?
yes, that would be great!
I spent a moment trying to make sense of this infomercial...
Do I guess correctly that when you write "all in ComfyUI", you mean "all in ComfyCloud using paying cloud based services that have nothing to do with ComfyUI" ?
SeedVR2, FlashVSR, Wan2.2 Upscale, Fast GAN, and also traditional non-AI upscale workflows are all open-sourced and available locally! We also attached all the workflows at the end of the article. You can always download the workflow and run locally!
Sure, but you didnt attach downloadable workflows at the end of the article : you gave links to workflows on ComfyCloud. Whenever I click on the links, I get a message asking me to choose a paying plan. That does not sound very "open source" to me, even if the upsacle models are effectively downloadable elsewhere, if you manage to get the information from other sources !
An by the way : I really do appreciate the great open source work the ComfyUI team is doing... I was just a bit disappointed to find that you were pushing proprietary solutions and did not give easy access to the WFs.
Thanks for the feedback. All the template workflows are open sourced as well, and you can find them in our repository: https://github.com/Comfy-Org/workflow_templates/tree/main/templates
We didn't link the workflows directly since some of them do require custom nodes and it might be confusing for people who don't have them locally already. We thought that it will be helpful if users can take a look at what the complete workflows look like and then they can decide how to use them.
Thanx for your answer, Jo, I was not aware the WFs were separately available.
"Video Upscale: Fast GAN" actually uses an ancient (Real)ESRGAN model which is very slow compared to faster and newer GAN upscalers...
That's good to know! Which newer GAN upscaler are you using recently? Maybe we can update the workflows with our creators!
Really good article, is all this available on comfy cloud?
Yes all of them are available on comfy cloud. The links at the end of the article will take you direct to the workflows on cloud. You can also choose to download the workflows if you prefer to use some of them locally.
Ty I will try it out!
I find Topaz is poor for illustration, smooths off the character of any natural media style. I find a tiled upscale with appropriate loras will do a better job for very high res.
Oh that's good to know! Would you mind point us to the tiled upscale workflow or the model it uses? Is it open source?
I put it here... https://discord.com/channels/1218270712402415686/1451352291880997026
the key to tiled upscales is to be able to prompt and denoise individual tiles . Ultimate upscale doesn't allow this so you get hallucinations in tiles where nothing much needs to change, a sky for example. I've done up to 16,000 pixels across but you need to add more tiles so that each tile is the ideal size for generating, IE about one mega Pixel for Flux.