Відео

Links in the description, but here you go. civitai.com/models/1156226/3d-or-live-view-comfyui-workflow

In Forge or A1111 there is a script called XYZ plot. Replace SEED with CHECKPOINT. And select all the checkpoints you want to test. It will process each checkpoint one by one with your prompt and a fixed seed and produce a single wide image with all the image and labeled checkpoints. I use this on a 1K image in the img2img tab with a denoise of something like 0.5. What you want to look for are the changes you're after. Does the image contain the style and elements you're after. You can do this for Sampler types, and denoising, anything were you need some objective comparison. You'll never find the ultimate checkpoint unless you build one yourself. But find one that works for that project and stick with it for a while.. Eventually you'll have 5 or so you experiment with, buts not an exact science, but nothing about AI seems to be.

32K is half a billion pixels or 500 mega pixels. If a client asks you to make that, you can push back and ask why. Unless you want to have some type of interactive zooming digital art piece, then maybe, but there's no print application that would ever take advantage of that, even on a closeup billboard 10 feet high. And that's speaking from experience. Assuming you still want to produce a 32K image. You can produce it with this technique! Even on the average GPU. You just have to keep isolating the upscale image into manageable crops, and stitch the final image together in post. Eventually however, large smooth areas will hallucinate and start to produce weird freakish imagery in the sky or smooth walls for example. My suggestion is the upscale as I've shown to 10-12K using several upscale steps at using a multiplier that scales each time by approx 2.5K ... 2K to 4K to 7K to 10K (approx) after that you should focus on upscaling areas were you need further detail. Not the sky, water or smooth areas, you can but those have to be at a different denoise like 0.25. Now that you've got me thinking about this, email me and I'll help you further, I love a good challenge!

@@matthallettai Thanks for your reply. I'm very familiar with working with very large multi-gigabyte files. I do it on a daily basis. I am producing very large murals. 10+ foot tall at 150 dpi that are being viewed at less then an arm length. I have custom scanning cameras that can scan much larger than 32K. Just trying to figure out how far we can push AI generated images and what hardware is required to do it.

@@matthallettai Yes, it seems like large smooth areas become problematic.

You'd be the exception then! I have to mention it, so many clients expect high resolutions but don't know why. If you can handle 32K files comfortable you have all thats required. The only consideration is your GPU, but you can get around that by using GPU's online by minute. I'm interested in this, email me if you'd like help outside of UA-cam comments. Just find my website in the details, my email on the contact page.

@@matthallettai Thanks. I'll send you an email.

Hey, This is Matt, but using my personal account. I don't cover much Comfy, and even after all this time, I still don't use it if I can accomplish the same thing with Forge. I've made the most insane Comfy workflows, and when I go to revisit them I can never remember all the little adjustments. You can easily enough replicate this with Comfy, you just have to install the Ultimate Upscaler node, and there's a few workflows out there for free that do the same thing I present. You can also email me and I'll send you my Comfy workflow collection.

@matthallett4126 Cheer's Matt. I work in house at an arch practice as a visualiser and they have pushed comfy UI (we had to license it) so no access to automatic1111. I think it would be much simpler to use that instead,but no one listens to me, anyway cheer's for the reply. I'll reach out over Christmas via email 👍

Clear your huggingface cache. Its on C:\Users somewhere And delete that sub directory in \custom_nodes\ and start over. Happens all the time!

Thank you! I'm going to make a new video showing my custom night conversion script if you want to subscribe and see then when its ready. It will be free to use.

​@@matthallettai I was wondering how you learned loRa. Do you make a video of this as well?

There are many tutorials how to train a LoRA. I use OneTrainer, but you can use whatever is new. You can start with CivitAI, they have a really easy online method for training SDXL and FLUX models. Start with something simple, don't try using a 1000 images like I have, it will melt your brain and GPU.

Thanks man! Appreciate that.. Right now the best image to video is Runway, Kling and Minimax. Which are all online services. For local Cog is very popular, LTX just came out, and its a year old now, but Stable Video Diffusion. If you don't like the online generators, you won't like the local stuff! There's not frame to frame methods as of now that I know of.

You can delete the entire program from \Data\Packages\ That will work.

There are a few solutions for AMD GPUs. Each package has a command line you need to add.. Stable Diffusion and many other AI architecture has been built for us with the CUDA platform developed by Nvidia. AMD GPUS are great affordable gaming cards, I use to only buy AMD GPUs.. Still only buy AMD CPUs. Its too bad they're not useful for the AI tech that's come out in the past few years. Thats why Nvidia stock went through the roof btw.

Oh there could be so many reasons. Check the logs in the Comfy Tyflow dialog box. Can't remember were that is at the moment. Make sure it can "see" the depth of the model and your resolution,.. model checkpoint. Also try installing Comfy or A1111 on its own and test your Stable Diffusion workflow outside of Max.

Not sure what you mean, but I have other video and there are other videos on UA-cam about using SD to edit your renderings in post

@@matthallettai i mean that can i render a photo and put it in ty diffusion and make it to generate just grass only for example

@@ibdesign8207 I would only use TyDiffusion for 3D models you want to quick turn into an AI image. What you're talking about is inpainting. Its best to take that image outside of 3ds max and use dedicated tools for AI photo editing like Forge, Comfy, Fooocus, Invoke. etc.

Glad you found it useful!

So nice of you

That error means your controlnets are mismatched. You have to us SDXL controlnets with SDXL checkpoints,

Cheers!

Just found something I think works, I'll post it on the Facebook group so I can include visuals. Assuming you're there as well. If not execute the 'run_ComfyUI.bat' in the root directory of the Tyflow install. In the CMD window look for the host copy and paste it to your web browser. it looks like this... http. 127 0 0 1 : 8188 but I can't paste it into the comments.

Thanks Alessio! My production quality could be greatly improved, but I do my own editing, and by the time I get to that stage, I'm over it :)

I've tried that using before in A1111 using an output ""object color" element like you're describing and removing the preprocessor in controlnet with the Segment model. It kind of works, this was last year and the model was shit at straight clean lines and small detail, so in a case like this were the subjects are far away it won't help. I was testing it making a kitchen and controlling single elements like floors and cabinets with a segment mask outputed from Max. I've always come back to using some type of base image with the colors I want in img2img and adjusting the denoise like I did here. But things change so quickly and often it comes down to the quality of the controlnet model. When flux has a segment model I'll try it again, Flux is amazing at straight and clean lines.

Thanks! Glad you liked it.

I think you're talking about controlnet? Instead of generating a Zdepth pass, Stable Diffusion can make one with a preprocessor and a Depth model. Its actually the most common method when using AI since most users don't start from a rendering program. Just turn controlnet on, selected Depth and the defaults should work. You need a SDXL controlnet for SDXL checkpoints. Thats the quick youtube comment reply, it can get a lot more complicated.

Yes, you can make a batch filled with images and use a generic prompt and a denoise like 0.55. There's a batch node in Comfy somewhere, but last time I tried to find it, I could figure out which one. I mostly use Forge, its batch generation is dead simple. If you can find my email address email me and send a message,maybe we can figure it out together. It also sounds like you could be asking about IPAdapter, which is another beast.

Depends on your app! This is just Matrix, its an app manager for image generators.

It works like any other channel, just click the "Show Shaded Material in Viewport" make sure you have your bitmap and your UVW set to 2.

Thanks man! I appreciate the comment.

Yes, we can do that now. Remember we tried last year with your 3D characters. There's easier tools to achieve this now without a Lora for each character.

Only works for me with Chrome, and its because I have all popups blockers disabled. Try the regular Adetailer, not the Udetailer, I think its better now.

As I mentioned on facebook they're inspired frames. I used img2img to create a few frames because I couldn't get flux to create anything like what I wanted with just text to image. I was more interested in seeing how those images would animate with Runway.

Glad it was helpful!

You can download models required on Hugginface or Civit. Then you have to place them in the appropriate download directory. If you use a Local AI manager like SD Matrix, it has a download menu you can use for automatic placement. Best to start with a startup video for Forge or Automatic1111

You're welcome!

Its a late reply, hopefully you have it figured out. If all your packages get that error then its a problem with Matrix or Windows security perhaps. IF its just one package, that happens during some updates, there's usually a fix posted quickly on Github. Or you have to revert back to a previous version.

Thank you!

I extracted frames from some movies and built a Flux LoRA, the base model of flux has very limited WWII data.

Thanks but its Tyson Ibele whos' the genius!

There are AI addons for Blender, but I dont know if it has this cool unwrap feature.