ComfyUI EP07: How to use LoRA in ComfyUI [Stable Diffusion]

[Music] hello welcome to episode 7 of config UI tutorial in this episode I will talk about Laura which is the model that can help you generate some specific character some specific Clause that model cannot do by itself you can download the Aurora from Civic AI if you search in the series Ai and filter for Laura and for the best module that you want and then you will find many Laura that you can download some Laura is the style of the picture some Laura is the character some Laura is the clothing you can also search or click in this tab okay for now I will go to the character Aurora and I will try to uh use the character of this Bulma from Dragon Ball okay if you click in the Roman character Laura you will see this page and I will provide a link in the description and you go to download this Aurora you click download okay and then you save in the configure I folder yourself in the configure I model Laura and save this file okay in the lower page in the Civic AI the author of Laura will provide you some help that help to achieve the best result for that specific ruler for example some roller has the trigger words okay you can click to copy the trigger but that can here generate that character okay for now I will go back to the config UI and I will put the prompt to be the Dragon Ball bomber okay without any roller at first this workflow does not have any roller and if i q prompt I will get this final result which is mean that my model know that the Boma from Dragon Ball has blue or green hair okay but she does not look like Bulma yet okay to to be able to achieve the best result you have to load the Aurora which you can double click and search for the roller no okay you will find Lola Road and this note you have to input the module and clip and then output module and clip again okay for the model you you drag the model from the logic point and also the clip okay and then you just output the model to the clear sampler and the clip to the clip test in code both positive and both negative so that it means that the Laura help modify the checkpoint to be able to generate some result that that roller can do okay in this case this Aurora can generate Bowman from Dragon Ball okay I will not change the problem yet I will generate the result for now and look for the the result that if I have this Laura will the result change okay you hit heel prompt and this roller will take the effect of the output okay you will see that her face and her hair style is look more like Bulma okay however to achieve the best result you have to look for the some trigger one okay in this case this trigger word can help generate this in this kind of hairstyle and clothing so you click at this clickable and you paste okay yes instead of the the first one this one is some a special keyword okay I think this one is maybe it's Bloomer ponytail okay the author of this ruler maybe think of this keyword for this hairstyle okay and maybe maybe this this is another keyword okay and this is another keyword okay so you can use this keyword to to change the result for now I just copy paste the trigger word and try to kill prompt okay and now you will have the result that look resembled to Roma from Dragon Ball however if you take a look at the result is some kind of a strange color or some kind of noise this because the strength of the Aurora is too high it depends on the author of the roller in this case this Laura suggests that you should use the width of the roller above 0.6 or 0.7 is the best okay so I will try to reduce the weight of the roller in the configure end you can adjust the strength of the Run both model and clip okay this two number does not have to be the same you can adjust them separately however to make it as simple as possible I will use the same number which is 0.7 if you change the strength and then kill prompt again you will see that the result will change okay a little bit however the result this time is not strange anymore okay the noise is okay and and the the texture is okay now in this example I use the same number 0.7 for both straight of the model and strength of the clip next I will change the number if I change some number maybe the clip to be 0.5 until prompt again the result will change the results will change if I change the clip more than 0.7 maybe 0.9 the result will also change okay so even this number change a little bit the resolution if I put the script to to be 0.7 but I change the strength of the model maybe zero price it the result will also change Okay so so it means that even a little bit of this number change the result can change so if you want to play about this number to see which one provides you the best result you can use another special note that that can help you change this number okay I will introduce you the special note which you can download in the manager you go to the manager you go to install custom node and you search for the efficiency okay Fe us find this efficiency notes for confi this node can help you a lot in many ways I will show you okay but you install it first okay after finish installing your restart from vui okay it's not configured okay after you finish installing the efficiency node you can add efficacy node and go to the loader and you'll find efficient loader okay this one is the combination of many separate default node okay you will see that it has the checkpoint loader it has the Laura loader okay it has positive and negative problems it has the latent image size okay so it combines many stuff in this case I will copy the from of positive and negative here okay and let's change the right to be 768 with the same okay and after that I add the efficient node sampling and here sampler efficient this one decide to be work with this efficient loader okay so you just connect the stuff the Sorting is the same okay and you can just connect it okay and I will change the C to be the same as the top one and also the step also the CFG okay to be exactly the same okay sampler name sampler name and caress okay and this node is already combined the vaed code okay so uh the the result can output as the image so you just drag and save image okay for now uh okay you can load Laura here this is Lola name okay so you changed and 0.7 0.7 okay I will try to set exactly the same okay and I just you prompt and look at the result in the top one also generates the image because it's another flow okay and you see it's the same okay the result is the same so I don't need this note anymore down control and drag and delete okay and then drag this and move up without that so the many confusing notes is gone the result is just this three node okay so it's called efficient note because it's efficient it makes the note number of node less and efficient okay because it's combined many stuff together this is the first benefit of efficient loader and the second benefit of efficient loader is to be able to use X Y plot so you double click and search for if y you will see the Sy plus you click the X Y plot is the stuff that can generate multiple image that can change some value in the x-axis and y-axis Sy plot needs dependencies so you drag these dependencies from efficient loader and I'll put the script and you connect the script okay if you want to use XY plus you have to change sample state to be the script okay this one is very important and to be able to generate multiple image you have to specify the x or y of the XY plot for example if the X address out and search you can choose many input that you want for example I want to change the sampler of the result maybe first sampler is Euler the second sampler is maybe DPM plus plus 2m is the egpu this one is the same okay and the final sampler maybe I want to test the the latest one demon plus plus 3M sde GPU if you don't have this setting you have to update complete UI okay if you connect only if okay this one you you can select more sampler but I will choose only three if you already connected only X and you keep from it will generate these three image okay one by one so okay the result of this three sampler is this it will contain three image okay and if you connect the Y input here you search and maybe you change other stuff such as the step okay if I want to change the step maybe I want to start from 30 to 50 step okay and let's cut three and then Q Prime again now the result should be Knight image okay the result is nine image you can click to see it this one and this is the crosstalk result and you can choose which which one is the best for you and in this example I change the sampler and the step and of course you can change it to be the lowest strength okay to test the lower strength you can drag X or Y is depends on you in this case I drag Y and let's search for the Laura if my input Laura okay and it will tell you to input the dollar that you want okay you can choose mode between dollar name brand name and wage and all that bad in this case I will change Lauren names and weight okay and you can specify three set of Laura okay because this Aura count is three so you choose the rural that you want you can choose the different one is up to you but in this case I will choose the Boma and I will change the strength okay of the model and the clip in the value that I want okay maybe the first case is 0.7 both okay the next one is 0.55 and 0.7 and this 0.7 and 0.5 to play with many value and you can add more okay you can add more okay if you change the number you you you have to choose this again to to be able to specify both model and clip strength okay and maybe this 0.8 and this 0.2 is 0.2 and this 0.8 is is up to you okay and just generate the image and it will create five image based on this setting um foreign of the model and clip okay the first one is the module and the second number is the clip and you can also combine with another X but in this case I think you can understand the concept of this XY plot so it can help you generate multiple image with different setting okay to test which setting uh satisfies you the most okay and I'm almost forgot to tell you that this efficient loader can load multiple Laura and multiple control knit at once okay with this roller stack if you drag out you can choose Laura stacker and you can choose many roller okay and change simple mode or advanced mode in the simple mode that means that the weight of the Aurora is both for model and clip and in the advanced mode it's very separate okay whatever clip uh for example if I revert back to the to the output that we create okay maybe this one this one this one that we have both original flow and efficient loader okay and download more Laura I will try this one and add hand pause okay this this one is also Aurora it's already can enable you to create just pause without using any control net okay and I will download this one and Save in the Laura folder and I have to use this trigger but okay not hand okay and go back to confi UI and I refresh so I will use this roller stack okay and I will show first runner to be Uma and which which 0.57 okay and then the second Lola will be the hot hand okay maybe 0.8 if I use this Aura stack I don't have tools use rather in this Edition loader I can use Dollar in here Dora stacker okay and I have to put the trigger in the prompt maybe I put it up in yeah okay and for the original one if you don't use the Laura stacker you have to create another round note okay I will copy paste and this one will refresh and say hot hand okay and I will wait 0.8 I add and I have to share this note together so model and clip the output model here and I'll put clip here okay so it's it's look like you have to share multiple learn applicator but if you use this roller stacker you just stack okay and I will test the result copy from to be exactly the same and your prompt you see now our problem can make the heart shape hand okay because of this roller hat hand so you can choose I think the last stacker is more convenient okay because if you use the original one you have to shade multiple one together and you if you want to use more than two lower you have to change more but if with this raw stacker you just choose you can change the roller coat here okay it's very convenient so if you think this clip is useful for you please like And subscribe my channel thank you and see you again bye foreign [Music] [Music] foreign [Music]

Channel: AI Angel Gallery

Views: 2,694

Keywords: ai, stable diffusion, comfyui, tutorial, ai art, digital art, generative ai, สอน, เอไอ, artificial intelligence, sd, lora, character, style, efficient nodes

Id: VFl1VFCrB5M

Length: 22min 42sec (1362 seconds)

Published: Mon Sep 04 2023