L7: Live Model Merge in ComfyUI - Comfy Academy

[@amorgan5844]

This deserves so many more views! So clear and well explained

[@XuRZaL]

Thank you, this has been so helpful. ComfyUI is very user friendly once you get past the learning curve (which you have shown me), and you have absolutely made that curve so much shorter. On to L8!!!

[@RichGallant]

Thanks, the explanation of the ratio setting was very helpful

[@Mootai1]

Very interesting ! thanks ! There is still one think i don't really understand : Why do we have to use both KSamplers one after the other ? will it not work with one one ? Thanks anyway. It opens the door to a whole bunch of new possibilities.

[@Friedrich713]

I wonder that as well!

[@jasoa]

I was wondering the same thing. I can see doing it if you're switching to a third model for refinement or upscaling.

[@hindihits9260]

second k-sampler is used to refine the image, think of it as like generating the base image and then doing an img2img with low denoising in auto1111. although in comfyui, we have far more control like choosing different scheduler in this case for better quality

[@LaminarRainbow]

W00T! Thank you for your amazing tutorials Olivio!! 🙏

[@eddieyu80]

What if models are for different versions separately? What to do otherwise? For example, SD 1.5 and XL respectively.

[@ShogoKawada123]

Thanks for clearing up which "direction" the ratios go in for ModelMergeSimple, I couldn't find that info anywhere else

[@joeyc666]

Could you say more about what the CLIP merge does?

[@romulotemigue]

an if I wanted to load a few loras with this? Where do I put it? Can you make a video?

[@ysy69]

thanks, this can be very powerful

[@DelioPera]

Hey, Olivio. What GPU do you have?

[@Ace77t]

but how do you save it as a brand new checkpoint? i tried to add the CheckpointSave node but it does not actually save it, nevermind i found it they were in the Output folder

[@MatthewHardyMusical]

The revAnimated checkpoint keeps making blurry images. What's that about?

[@rakly347]

Are you sick? Get well soon!

[@AkoZoom]

lately but if ? .. is there a way to make this with SD1.5 models merged with SDXL .. for sure may be no.. but then ? to mutate the SD1.5 to SDXL ? may be YES ? as the refiner possible path ? how to make that in comfyUI ?

[@MrMorvar]

Is it somehow possible to merge SD1.5 and SDXL models with this?

[@OlivioSarikas]

no because they are not compatible with each other. but you can render a image first in one and then in the other by decoding it to a pixel image and then encoding it again to laten image with the second model you want to use

[@MrMorvar]

@@OlivioSarikas Alright. Thank you very much!

[@maciejbala477]

yeah it's not possible because CLIP for 1.5 and XL is different i assume

[@manuelR78]

Hey olivio, könntest du Mal bitte ein Video machen in dem du auf Sampler und scheduler eingehst? Klar kann experimentiert immer da umher mit verschiedenen settings, aber sicher hast du einiges dazu zu sagen in welchem Fall welches setting sinnvoll ist. Danke dir 😊

[@Veselin_Angelov]

I, A1111 simpleton, don't know which pack are the live merge nodes from? Also, is the second K-sampler a sort of a refiner?

[@Elwaves2925]

If you have the manager installed, you should be able to get the nodes from there by clicking 'Install missing nodes'. Most of the time it works that way.

[@ImmacHn]

The second KSampler is just that, a second KSampler, what this is doing in this case is taking 60 steps and dividing them into two complete denoises, which in my experience is overkill especially on low CFG, If you try this workflow I suggest you leave the first KSampler at 15 steps 0.7 denoise, and finish denoising with he second one, 20~25 of dpmpp2+karras should be enough 1.0 denoising. I think you can even go lower on steps but I'm not near a PC to confirm right now.

[@Empowernity]

can you do about IP Adapter Unfold Batch

[@sirjcbcbsh]

how to apply Lora with it

[@RodgerE2472]

You da man!

[@Skettalee]

Hey man, I hear a huge amount of noise across the overall video, checked if its my system but its only playing on your videos. It might not be important enough to reprocess the audio and reupload the video but just something I wanted to mark down just for your knowledge. I wait for your videos like im waiting for a saturday morning cartoon as a kid!!!

[@PADATWO]

A model is a lora ?

[@chrishillery]

No, a model is a checkpoint.

[@maleeshamevan]

thank youuu..//

[@gnoel5722]

Thanks for the video. I don't really understand the point of the clip merge, I understand that if I ask for a "cowboy" in the text prompt it won't look the same in both model but wouldn't that be the ModelMergeSimple node job? What I don't understand is, what are you merging if the "clip prompt" for both model are the same? Maybe I don't understand what the CLIP exit pin really is... I would appreciate if someone can explain what merging the CLIP from both model and playing with the ratio will actually do.

[@koen3624]

Hi Olivio, do you have more information on selling comfyui workflows to advertising agencies?

[@blisterfingers8169]

lol