Animatediff LCM Lora in ComfyUI for Faster Render Times and Superior Results

[Music] I spent some time testing out the LCM Laura to speed up my rendering time using animate I created three different workflows following a few different tutorials hoping to achieve better results even though the time generation for these workflows were impressively fast from the LCM Laura the results were very very terrible to fix this I will explain the benefits of animate LCM Laura and show how to correctly create the notes using the proper settings and animate them to achieve faster Generations better results and high video quality all at once even without using an upscaler to get started first visit the comfy UI animate LCM page once we scroll down the page uh we'll see the installation section to see the instructions provided I'll follow the link here to haging phase we will download this two files which is the animate LCM Laura and also the checkpoint to download the first model click on the download icon and place this into the cfy UI directory models folder and the animate diff models secondly we go for the animate LCM Laura file which we also click download we save it into the cfy UI directory under the models folder lores and we have it saved I already have mine here before starting go into your custom noes and make sure these extensions are installed so let's install the cfy UI sampler LCM alternative this will allow us to use the LCM sampling methods to generate the animation go into comfy UI install custom nodes search for LCM choose the comfy UI sampling LCM by joke cat click instore and be patient once this is done additionally search for advanced clip text and code we will use this to replace our normal text clip Noe and I also click install and be patient close the manager once all of this is done click update all to make sure you have the latest version of animate also update C UI and restart the server so I'll load my workflow which I previously tested with the three workflows that failed next to the three workflows we already have I'll build the ultimate workflow here for easy comparison to the field results so I'll also right click edit group and I'll change the color here to Red for our first node I will double click and search for the load checkpoint in here I would also search for the Laura as our second node to use after the low checkpoint now from here I'll drag out the Laura to create the text and code I'll duplicate this as our positive and negative prompt and also I'll change the color to green and red um for easy recognition drag out the clip to the negative node the model from the checkpoint to the Laura as well as the clip to the Laura next we need a k sampler but we will use a different node here which is the sampler custom node this is a personalized sampler built into comf UI that comes along with the new extension we just installed a similar to K sampler I would connect the proc positive to the positive and also the negative to the negative so I'll drag out the latent image for our video frame size which will be 512x 512 and the number of frames I'll go for 120 so to see our video generation I'll drag out the Deno output into a VA de code node I'll scroll up here to get one of the VHS nodes we already have to keep the same frame settings I'll change this to red and then I can connect the vaed code into the VHS node I would also drag out the vae to select the load vae node for the animation so I would select the same checkpoint to get the same results from all the other workflows so I'll delete the load vae node and rather connect the vae from the checkpoint and I'll reroute this all the way um into the VA Noe so from here we now have our simple text to image workflow completed uh so let's keep let's keep things tidy and make space above so we are going to build an animate workflow together with the LCM sampler extension now double click and search for load animate diff model search for the next model which will be the apply animate diff model connect both of these noes into the motion module of the apply node so to go for the third node I'll just search for use evolve sampling node for animate drag to connect this into the motion modules of the evolv sampling node drag out the context option and here I'll select the cont text option standard uniform this will be already connected and I'll move this to the right so we are almost done and we need to include our first LCM node so I'll just double click to search for LCM and I'll select the first one which is the LCM schedular node connect the evolve samping into the LCM Schuler select everything here and I'll change the color to Yellow for us to just keep track of everything uh so from here let's connect animate St to the text prompt workflow and uh this is very simple actually drag out the model output from the Laura to the evolved sampling node animates the then the evolved sampling node into the sampler custom which goes into the input called the sigmas uh next drag out the sampler for the custom sampler and in here I'll just search for LCM cycle and select this one here so this is going to help us with the value of steps which is used in the cycle and can be adjusted based on our generation settings so now let's input the right settings into each node to get the correct results inside the Laura we are supposed to use the animate LCM Laura which we downloaded earlier next under animate div we are also supposed to choose the animate LCM checkpoint which we also downloaded earlier the combination of these two are very very essential also change this to the LCM 100 and going back to the page I selected this as my option to use let's move to the sampler custom to change the settings here as well so I'll change this to fixed and also for the CFG we are supposed to keep this between 1 and two I would also change the SE to be 100 at fixed the empty latent image I'll keep everything at default and I'll change the eura steps to two for the sampler LCM cycle so scrolling up I'm going to replace the text noes so I'll search for text scroll down to find the clip text en code Advanced the settings here needs to be mean and also this needs to be a111 so I'll move this up here and connect the clip from the Laura into this node which will be our positive and also I'll connect this to the sampler custom node delete the positive pront we have already and change this as green and also I'll duplicate this as our negative and I'll change the color to Red also connect the clip from the Laura to this and also connect it back to the negative into the sampler custom node so from here we are done to see a completed workflow for animate including the sampler LCM noes and also the LCM models so let's put in a prompt into our new workflow the ultimate workflow and I'll copy this from one of the workflows here and I'll just drag down to paste this into the new animate workflow using the LCM sampling schedules I'll also paste in the negative prompt as well all is ready to go from here I'll just kit Q prompt to see the generation of the new workflow compared to the previous three workflows I already have and man this is actually mindblowing comparing all the results next to each other so my realization is that the other workflows are not using the sampler LCM nodes so even though they still included the LCM models which we downloaded the results are still poor without including the sampler LCM nodes along with animate so um Let's test a different workflow let's give this a second test and we are going to focus on the speed for the different workflows we have now and also the best results from the four different workflows since this was extremely fast increase the frames from 32 to 120 for all of the test workflows so uh from here I'm just going to hit Q prompts to see the results for our second test generation uh to save us some time here I'll just skip straight to how this turns out all right so uh man this is a huge difference compared to the other workflows on the left which was not built the right way uh you won't even believe this was generated with LCM in addition uh the quality is Flawless using the animate LCM workflow so this previous example was also 120 frames and we can see the time executed here was only in seconds which I just changed the checkpoint to use a Disney model style so using the proper animate LCM workflow you may even create a straight one minute video in a matter of minutes with faster and still better results all at the same time so the most question I have received from you guys is uh the GPU I am using for my animation and with this formula regardless of the kind of GPU or vram you have you can reduce your animation waiting time by taking advantage of the animate LCM Laura and without the use of note skilles you can also check out this tutorial still of attive inside comfy UI to create the right workflows for your project smash the thumbs up if you found all of this exciting to speed up your workflow and I'll see you guys in the next video

Channel: goshnii AI

Views: 986

Keywords: Animatediff LCM Lora, LCM, comfyui, stable diffusion, Stable diffusion tutorial, animatediff, animatediff comfyui, animatediff tutorial, Animatediff lcm, animatediff low vram, comfyui animatediff, animatediff lora, animatediff & lcm lora, animatediff with lcm lora comfyui, lcm lora comfyui animatediff, animatediff with lcm lora, lcm lora animatediff, lcm lora, comfyui tutorial, stable diffusion animation, stable diffusion tutorial, lcm, ai animation, ai, text to video

Id: mYmopXiFW2I

Length: 9min 45sec (585 seconds)

Published: Thu Feb 29 2024