What you see in the video is not Magnific. It's my AP Workflow for ComfyUI (which is free).

@ no I know I mean if you know another paid program can you suggest?

@@canberkguitar-sg6qu No, sorry. I don't use any as they are not competitive with SUPIR.

Thanks for your kind words. Unfortunately, for me to host APW on any ComfyUI hosting provider, I must have an agreement with the company. It's too onerous for me to maintain the workflow for free, and I don't want a 3rd party to distribute a workflow of mine if it's not working properly. On top of this, APW 12.0 release will offer a huge number of new features. Hosting APW 11.0 doesn't make sense.

You are the second person to say that :D Probably, this is WAY LESS creepy than what I wanted to generate, so I didn't realise the impact it already has. Thanks for stopping by and giving your feedback!

@@peacetoall1858 APW can run with any amount of VRAM. By default, APW is configured to look for certain AI models in certain paths (e.g., FLUX 1 Dev FP16) and some of those won’t run on a 4GB VRAM configuration. However, you are free (and encouraged) to replace those settings by pointing to AI models that work with your system and you have downloaded yourself (e.g, FLUX 1 Dev FP8). What AI models you use with APW is entirely up to you.

@@Perilli That's cool. Thanks!

@@JackKnifel sure. SUPIR is part of my AP Workflow for ComfyUI. You can find it here: perilli.com/ai/comfyui/

SUPIR easily beats Topaz in my tests (see a couple of videos I published about this), but you are right: it's very resource intensive. I have at least one member in my Discord server who has your same configuration (12GB) and he reports that, in the past, he could upscale with SUPIR up to 4K without too much trouble. But recently, he had more and more issues and now he can't go past 2K. So I am telling you the same thing I told him: It's possible that a combination of sub-optimal parameters + a suboptimal software configuration (the version of your Python packages in your ComfyUI venv, your version of Pytorch, your ComfyUI launch flags, etc.) is disrupting SUPIR. But the good news is: if he managed to use SUPIR with 12GB VRAM, you can, too.

Don't feel bad. Even tech experts who read my newsletter for 2 years still hate the idea of sitting down and spending time crafting the perfect prompt.

@@tmode1960 First of all, thanks for your interest in APW. It's great that you managed to install everything without problems. ...but if you did, you already have the workflow. So you mean the web front end? If so: Check again the "How to Download APW" section of the documentation. It has two links to the two web front ends I created for APW 11.0. And yes, as I say in the video, one of the two is a token of gratitude just for Patreon supporters.

Thank you. Do you mean having the LoRA training in the web front end, or in the workflow itself?

The audio is unedited. Occasionally, NotebookLM repeats itself. And it hallucinates facts not in the newsletter, too.

First of all, thanks for your interest in supporting research. Sure you can join: www.patreon.com/Perilli Do you have any issues?

@@Perilli Every time I try to checkout, I get the message "You can’t join this creator’s membership." However, I'm a paying member for several other creators on Patreon, so I believe there is no issue with my credit card.

@@kp_mark That's odd. If you have Discord, please reach out to @perilli there. I'll try to figure it out.

Without altering the subject or background of the raw photograph, all I want to do is process it by transferring its style, colour, etc. from another edited image.

Focus on the ControlNet and IPAdapter functions of AP Workflow. I'll publish videos about this, too, at some point in the future.

@@Perilli thanks I will try it out

Thank you for your kind words. You'll be please to know that I just released AP Workflow 10.0 to the general public as a free download: perilli.com/ai/comfyui

Just download the AP Workflow 9.0 (or just look at the image of the workflow with all the settings of SUPIR) here: perilli.com/ai/comfyui/

Go to perilli.com/ai/comfyui/ and then join the Early Access program via Patreon or Ko-fi. Or wait for the public release which will be free.

The workflow I am using is the one in the URL visible in the video. You can find it here: perilli.com/ai/comfyui/ But yes, I'll publish tutorials when I release AP Workflow 10.

The workflow I am using is the one in the URL visible in the video. You can find it here: perilli.com/ai/comfyui/

​@@Perilli Sí, yo entré a esa URL pero desde ahí no se puede abrir ningún programa, es pura propaganda de lo que se puede hacer pero no te deja abrir nada para usarlo. Yes, I went to that URL but from there I cannot open any program, it is pure propaganda of what can be done but it does not let you open anything to use it.

The AP Workflow does not use Ultimate Upscaler but CCSR and SUPIR, two new upscalers that do a much better job. I'm thinking about doing videos, yes. When I'm ready, I'll publish one about this topic.

​@@Perilli I will pay for video tutorial from beginning step by step how to install and what to instal and have working this on my PC, anyway subscribing for future videos

Hi @vo.images. Thanks for your interest. Unfortunately, I don't have yet enough subscribers and views to unlock monetization on KZbin. So, initially, support videos like the one you mentioned will come via Patreon & Ko-fi. But they are coming.

SUPIR is a new technique to upscale images (both AI-generated and non). The SUPIR AI model is free and multiple people have implemented in various ways, in commercial apps and free apps. One way to use SUPIR for free is as a node in ComfyUI, which is a free automation engine for Stable Diffusion. While ComfyUI is free, it's also very difficult to setup and master. To help a bit with the second part of it, I create an automation workflow for it called AP Workflow. Which is free, too. In the upcoming version of AP Workflow, version 9.0, I'll provide the SUPIR node with a configuration that I think works well. For free. For people that cannot wait until the release of the AP Workflow 9.0, I created an Early Access program. You can sign up and you get access to the AP Workflow 9.0 Early Access pre-release today. Inside the newest pre-release, called AP Workflow 9.0 Early Access 3, you'll find the SUPIR node. The Early Access program is not free. And I provide no documentation for the pre-release. So it's meant for people that know what they are doing. If you want to know more, search for "Early Access" inside the official website of the AP Workflow: perilli.com/ai/comfyui/

@@Perilli Okay, but I did not understand the point. Is it a website or a program on Windows? Or on the phone?

By doing the exact opposite of what you mentioned in the previous comment. According to the researchers, a highly detailed description of the image helps increase the details during the upscale process. That's why they invoke a VLM as part of the process. However, it depends on how good the VLM is. The previous 4x test I published leveraged an automated captioning produced by Moondream (which is one of the main VLMs in the Caption Generator function of the AP Workflow) as positive prompt for the SUPIR node. That caption, which was very detailed and accurate, generated a much worse result as the beard was confused as cotton, or otherwise soft material. In this replacement 4x test, I used a manual, very minimal positive prompt and I got all three beards that look like beards. The whole process is less intuitive than it seems, and I'm still doing a lot of experimentation to understand if this result can be improved with a better prompt strategy or if the model performs worse with source images that already are high res.

​@@Perilli That's odd. For me it's the opposite. I should note that I'm not describing the image in the positive prompt. I do it in the "prompt" field in the "General Options" panel. Maybe that's why you are getting worse results? In the "default positive prompt" and "default negative prompt" fields in the "Stage2 options" panel I'm just putting generic keywords like "cinematic, maximum detail, skin pore detailing, etc", and for the negative prompt "painting, illustration, cartoon, jpeg artifacts, etc". I've never had images with poor detail doing this. BTW, I'm not using AP workflow, but Gradio. Where can I read the researchers' article on how to best use the upscaler, if you don't mind asking? I would love to learn a bit more about the best practices and how each setting affects the output.

@@bgtubber We are saying the same thing, even if it seems I'm describing a different location. The AP Workflow organizes SUPIR fields differently. The paper is here: arxiv.org/pdf/2401.13627.pdf but most things can be understood by reading the comments on the default repo and in the code. It's quite messy.

Yes and no. The upscaler in APW 8.0 is powered by a CCSR node, which can do extremely high fidelity upscale at big magnification levels (I tried up to 10x) even on consumer GPUs. But if you want a more creative upscaling, you must also active the APW 8.0 Image Enhancer function, which is very complicated to setup, but works with limited VRAM as well. This new SUPIR Upscaler is more creative than CCSR, so it's not good if you want extreme precision (for example in fashion photography), but it's infinitely easier to configure than the Image Enhancer of APW 8.0. The other problem is that it doesn't work too well, for now, at high resolutions. So I couldn't test it past 2x. And it requires a massive amount of VRAM. The idea for APW 9.0 is to combine CCSR (for use high-fidelity upscaling) with SUPIR (for creative upscaling) and get rid of the complexity in the Image Enhancer function. But we are still far from that. A few things must be developed. That's why I said "preliminary support".

@@Perilli I've spent days pushing old lowres midjourneyrenders and some of my own 3d-renderings through your AP8 pipeline and got mostly great results. But you are right - for some images I would have wanted a little more creativity. Looking forward for version 9.0

@@planethanz Out of the box, APW 8.0 Image Enhancer function is configured for high fidelity. So you can have more creative results for your images, but it's a bit hard to configure. Hopefully, SUPIR will make it easier.

That's exactly right.

Thank you, @oliviertorres8001 ! The AP Workflow orchestrates dozens of AI models. So yes, its installation is a bit of a challenge. 99% of the issues are related to the errors in the ComfyUI virtual environment, conflict with existing packages, missing models, or not adjusting the path of the models in your environment. The documentation has been updated this week with a few additional clarifications. Hopefully, that will help.

Thanks for your kind words @jankrejcir

Great question @strollnvibe. I'll reply in an upcoming video. Stay tuned.