Big thanks for this video! It really showcases the serious limitations of comfy as an interface for artists. It’s great tech, but it’s super limited. Thankfully, enterprising developers are building bridges to photoshop, so we can quickly build the composition then run the generation on the composition.
Really appreciate your videos. This spanish-speaking youtuber love your videos a lot too lol ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-VgE99qY2VSM.html
Hi there, this is old enough but I hope someone will point me in the right direction. When I try to use UltralitycsDetectorProvider I get a red cross of the SEGM detector. Can somebody help me with that?! Thanks a lot!
@Ferniclestix: RE: IPAdapters/Clip vision. The current version of IPAdapterPlus by Cubiq (Matt3o/Latent Vision) has a "Ipadapter Uniified Loader" node that takes the guesswork out of finding the right IPAdapter-ClipVision pairing.
caracter poses and full characters are possible but multiple character source images might be needed, you would want to create a full body character from a portrait, then use the full body charcter for full body poses. It takes more work but easy enough once you have the source images you need.
Consistent Characters is impossible till now using sd , all it weak tricks . might be soon who can developing real ipadapter for character likness and share them , otherwise atm all txt2img,img2img,cotrolnets scene has nothing call Consistent Characters , it's a single lie from many. keep the good work
yeah, I agree, this method is about as close as you can get using current tech. There are ways of improving even this but its more about building a big reference library of characters and poses to get the job done. lots of work.
Awesome video as always, glad you could make that one. Always a pleasure to get your approach around things like that. I hope you'll get better quickly ! Thanks too for the shout out on ComfyUI_MaraScott_Nodes - AnyBus :)
It would be helpful if you immediately showed the end product in the intro of your video so viewers can know if this is the tutorial theyre looking for. and find out of this covers video and not just stills.
I'm new to ComfyUI. I noticed that your KSampler & Main widget shows a live preview of each step of the image generation as it generates. How do I enable this?
To future readers: I figured it out. It's called "ComfyUI Manager" and is easy to install by git cloning the repo into the right directory. Guide is in the github repo page
Enjoyed the video. I think we all go through the same frustration with these programs. I was watching a similar Lora stacker tutorial and they used a CR apply lora stack node to connect all the bits. But I could not find it. You may have better luck. Here's the RU-vid video. video ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-z4UIFCNS_FI.html. It looks really good. But where's the CR apply lora stack node? It did not show up when I searched for it. Thanks for your video.
thanks for the cool video. how can I make the generation happen only after the picture changes? I turn on the auto queue and change flag and the generation goes on constantly, but I only need it after changes in the picture.
Quick thing i found out: "IMAGE" input/outputs only have R,G,B channels, while "MASKS" only works with ALPHA (or, at least, have a single channel). So "Load Image" has two outputs: "Image" (RGB) and "Mask" (A); it splits the channels that way. So if you try to convert from "Load Image" to mask as alpha you get an error. Following the reasoning, nodes as "Mix Color By Mask" use "image" as input (and not mask) so we have more freedom. Also they have only r, g, b options and not alpha because "image" data in ComfyUI does not have it, but on the other hand if they used "mask" they should use a mask (1 channel only) already processed previously
Problem is with xy plot, because it gets it dependencies directly from the loader and connected to the ksampler, it ignores everything between them, like controlnet or ipadapter. It just keeps generating images, as your ip, control stuff would be set to bypass :(
Yu could have quickly discovered an answer this way: Double click to add a node, use the *second* dropdown to filter by the type. select lora_stack and voila, you can see all the nodes that use lora_stack as an input or output type.
I was under the impression (and based on a lot of testing) that you cant actually put the trigger word in the prompt. Do you have something specially installed or special that needs to be done?
Short and to the point. I had been wondering why some of my Lora's weren't seeming to work - turns out they're just incompatible with the checkpoint 😅👍
