AI Tutorials since 2020. Learn Stable Diffusion, Generative AI, Large Language Models, AI Animation, Voice Cloning, Text to speech, VQGAN, StyleGAN, Style Transfer, Super resolution / upscaling, Inpainting, Outpainting and many other amazing AI tools. Focused mainly on Free Open Source Software.
Nerdy Rodent is an AI enthusiast with over 30 years professional computing experience. Make this channel your go-to for anything related to Artificial Intelligence to benefit from all that experience for free!
Don’t forget to like, share & subscribe.
For the best AI experience at home on your own PC I suggest: * Ubuntu 22.04 LTS * A modern Nvidia GPU (16+ GB VRAM) * 64+ GB RAM * 2+ TB SSD * CPU with 6+ cores
Nerdy Rodent - the OG AI tutor creating nerds from way before AI was suddenly cool 😉
Not suitable for children. Please use AI responsibly.
Hi Nerdy Rodent, just discovered your stuff, amazing! I'm getting an error when i try run the model. The error is: Error occurred when executing IPAdapterModelLoader: xxx missing 1 required positional argument: 'ipadapter_file'
How can i combine Harvest +Diffgrad and Adam , wenn i just use the RVC Interface from mangio or Applio ? So in the Json file in "Learning Rate i have these entry : "learning_rate": 1e-4 . Why is it different to yours ? How should i change it ?
how is everybody not stuck at the checkpoint? comfyui Prompt outputs failed validation CheckpointLoaderSimple: - Value not in list: ckpt_name: 'sd15/realisticVisionV51_v51VAE.safetensors' i downloaded the fabled stabe difussion one point five but it doesnt seem to count. what is the difference?
These new models do look pretty nice! I do find it odd that we end up dedicating more and more GBs to text understanding and seemingly smaller and smaller fraction of GBs to actual images... I wonder if this impacts how much we would be able to finetune these models. I also would not mind seeing some models going the other direction, terrible natural language understanding but cramped with image data, like a PonyXXL...
I've tried some of these tools and also experimented with building simple versions myself. I think for simple "handful of documents" scenario these can work quite well, but they usually only implement basic index search and context injection, so if you ask questions like "count the number of people mentioned" or any question where you need to look at multiple places in the document, they usually do not work well. You would need a pretty elaborate branching workflow and a switching process to choose processing branches to deal with complex queries, and these usually cannot be built without being designed specifically for a dataset. I think the easiest way to get a semi-decent QA system is probably to build one with Flowise, which allows you to compose LLM nodes similar to ComfyUI.
Thanks for the great detailed video - but I think we are getting overwhelmed with choice that doesn't now seem to offer a substantial 'reward'. In other words, not much difference between them really for all that extra time fiddling about with different settings :) I can see the use cases in video generation for extra speed - but far less so for static image generation.
Hi Rodent, I really like Pixart-Sigma. In fact, I love it! But I noticed there are also ControlNet checkpoints (and there's mention of lora support ) in the repository. I've never seen anyone come up with a way to use them. What say you?
Nice, but at the same time I'm scared because there are just new options to explore by trials and errors. So let's spend another half of a year generating images with different parameters then another half of a year for previewing them and picking. I'm afraid that some concept is lost in this nightmare. We needed a tool to make images quickly based on what we think. AI text interpretation and image generation was supposed to do that. But observing all the communication, tools, videos and discussions about that, I see some countless amount of hours spent worldwide on trying to deal with it's weaknesses. Of course AI/ML direction is desirable, but I believe the future is moving it into 3D domain because this is reflecting our world, doing some deep integration of AI and 3D engine, physics, collision detection etc. Instead of spending hundreds of hours on trying to fix AI artifact, maybe it's better to spend it on literally manual creation of some part of 3D model for Blender and then let AI to do arrangement of models in the scene etc. Combine 3D "thinking" and 2D like we have now in AI generators for backgrounds and textures. "Woman on grass" example of SD3 medium. If there is just customizable and parametrizable 3D model to be posed and placed in the surrounding environment by AI including rather rules of physics than analyzing billions of parameters on 2D images that are indirectly trying to reflect mapping from 3D world to 2D image, then I believe we could avoid "body horrors" and many other artifacts.
Thank you. I have a question. Is it possible to add HighRes-Fix Script to the Custom Sampler? I know that you can connect a second KSampler and then HighRes-Fix Script. But I'd like to be able to do it directly.
I have to imagine that less than 10% of your viewers actually use Linux. So, to say the Lumina install is easy (unless you use Windows) is really completely missing the point. Installing Lumina on Windows is not fun at all.
webui-forge and a1111 are great ways to discover more errors than you ever thought were possible, IF you can manage to even install them. and to make matters even more interesting, A1111 is being maintained and webui-forge is not. I've got a 12GB card and it's not enough even though their troubleshooting wiki has examples for 4GB card users. --lowvram nor --medvram worked, --xformers will never be installed, ipadapter v2 will never install, etc. Changing checkpoints will freeze the whole computer, just loads of learning experiences. No idea what they're thinking.
There was a bug in ComfyUI [CFG++ Samplers] The bug causing malfunction in the recently added 'euler_cfg_pp' and 'euler_ancestral_cfg_pp' samplers has been fixed. Update ComfyUI to latest version.
Glad they fixed it. After the year we've had, I guess it's only fair. In this case, when I say "fair", I mean it in a way where: "fair" is euler's day off. not funny? Euler? Euler?
I generally use AYS but for unsampling, gits can get away for 2x6 steps (6 unsampling, 6 sampling) at 1080*1920 without any diffusion lora and the image is nearly identical to a 20-30 step with a regular scheduler or 10 steps ays
idky, but the algorithm has been hating on you lately... i haven't been recommended one of your videos in "a rat's age". that's like "a dog's age" but nerdier xP
@@fureytha no **** sherlock... i've been subscribed for like 2 years... which is why it's weird i havent had a video show up my recommends for like a month
I hope to live in a world where Open Source lives on and blackboxes are still a thing. It's fun when life still has a little mystery. Especially science/tech.
Interesting. Which samplers are better? I'm a bit of a newbie. I've been using DPM++ 2M Karras and Euler/Euler Ancestral in ComfyUI with the regular KSampler node. Should I switch to using these new samplers and schedulers and what would the benefits be? Any speed or quality improvements? I couldn't understand from the video. I'm mostly doing img2img stuff with controlnets and IP-adapters rather than generating stuff from scratch. Would these benefit this use case?
@@Ethan_Fel Ok, I've just tested this. In my limited time trying it out, it looks like using GITSScheduler with DPMPP 2M vs not using it gives me 50-90% speedup for the same quality. I get similar speedup with AlignYourStepsScheduler too (AYS). Neat!
Could you share this workflow please? Your tutorials are exceptional (thank you). The reposer workflows on your drive don't match these videos. While there are new updates since your video, I still think yours hold a lot of relevant value. If you want to update the workflows, I wouldn't complain either!
Being such an old video, the nodes which exist then no longer work. I did, however, use both of your suggestions so that people have the option of using the newly updated version, or do any updating themselves! - github.com/nerdyrodent/AVeryComfyNerd?tab=readme-ov-file#list-of-workflows-available
@@NerdyRodent Thanks for your prompt reply. I can't seem to get the output to be the same as the original picture. They are close but very obviously different. Would you be so kind as to tweak my workflow so I can better understand what to tweak in the controlnet to get the desired results? I've been tweaking and adjusting variables. The only thing I've achieved is a headache.
@NerdyRodent Do you know if there is a way to Apply Powerpaint WITH A DENOISE LOWER THAN 1?? I cannot get PowerPaint to work with denoise different than 1.
@@NerdyRodent thank you for reply. I googled the error and couldn't find anyone with the same error. I uninstalled and reinstalled onnxruntime, used the .pt models instead of .onnx in the DWPoseEstimator (node 238), tried other SAM Loaders, SAM detectors (idk what they do), updated COMFIUI (portable) via manager and .bat (ran the .onnxruntime remove/install cycle again). Nothing.
@@NerdyRodent I’ve been suffering with these errors for 3 days, I can’t fix it in any way, maybe someone knows how to do this at the current time, where to get the right models and nodes? that would be incredibly cool!
nope, sd3 has better architecture but safety training as well as license are dogsh*. And you know these scores can be gamed. No way in hell dalle3 beat midjourney. I use all of them on a daily basis.
Thanks for the great overview! HunYuan is better at following most prompts you gave it, and better at composition, and was better at proper human faces and hands. But I sadly think its name will hold back its popularity. Its name doesn't sound cool. It sounds confusing. 😅 Edit: Oh it's by Tencent, now the quality makes sense. They are some of the best in the world at this stuff.
@@NerdyRodentOh yeah true, Tencent are great at creating IPAdapter stuff, which I hope helps the popularity of this model. I can't even remember the name right now to write it again in this comment. Yuanhun something? That's a problem for its popularity. 😅 There's only one or two fine-tunes for it right now. Really hoping it gets more popular soon.
@@NerdyRodentThe difficult name strikes again. I couldn't remember its name and literally had to check here again. I don't think the name was a good idea. 😁
Hey, I wanted to try out your workflow and after installing every model and every node it stops at the "Batch Creative Interpolation point. It says something along the lines of: ""ipa_weight"] for x in bin.weight_schedule] (...) in apply_ipadapter" Do you know what to do to fix this problem and begin animating images?
I was wondering, when do you think would be a good time to update to Ubuntu 2024.04? I'm a bit worried some programs don't work yet on it. Also, great video, as always! :D
The accent wasn't too bad but it was just going from region to region. Imagine someone giving a "British" accent and they say "gov'nah" with a royal accent... it was like that. hahahhaha