ComfyUI Workflow<ul><li><a target="_blank" rel="ugc" href="https://youtu.be/uAjcFdRi0cs">How to download, install and setup Comfy, custom nodes, install facelib, create symlinks...</a></li></ul><ul><li><a target="_blank" rel="ugc" href="https://www.youtube.com/watch?v=OVOb2qFT5U8">an excellent explanation for beginners. apart from the fact that the denoise value is described incorrectly, highly recommended. by jboogx</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/articles/4642/links-resources">Links / Resources</a></li><li><a target="_blank" rel="ugc" href="https://youtu.be/GCmAsFQz8lY">How to Add Custom Colors</a></li></ul><ul><li>navigate with the num keys 1-9 in this workflow.</li><li>every group that you dont want to use can be bypassed.<ul><li>e.g. if you want to load a video or images and do just a face swap on them... go for it.</li></ul></li><li>every sampler can be used as standalone, that means if you run the first one (e.g. KSampler (Denoise)) you can load the images created from this sampler via the input images group and run the second (wich can be every sampler in this line 1st Pass, 2nd Pass and or Latent Upscale) and so on.</li><li>the reason for that is, if you run out of memory on some point and you most likely will, you can load the last created images and re-start from there.</li><li>you can also use this as Post-Prod. (load a video or images and start at the face detailer, face swap, face restore or what ever comes to your mind) BUT bypass the groups, don't use mute (NEVER).</li></ul>Version 4.5<ul><li>noticed an issue with the length of the prompt, batch prompt raises tensor size errors if the prompt is too long so its possible to disable it now if not used. (the main reason for this early update)</li><li><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ccefaf21-97f8-41a9-bf11-6fd7f1e87f02/width=525/ccefaf21-97f8-41a9-bf11-6fd7f1e87f02.jpeg" />if you want to use batch prompt and this error apears, change you prompt.<ul><li>the description of the subject should be in the batch prompt not app-text.</li></ul></li></ul><ul><li>input area changed<ul><li>put multible files (videos) or folders (images) in the multiline text fields and select the line (0 to x) you want to work with. you can switch it to increment so it changes with every queue<ul><li>e.g. if you have splitted a video into scenes and saved the images into different folders you can run all folders with the same settings if you like.</li><li>or if you want to run multible videos with the current setup over night.</li></ul></li></ul></li><li>changed the checkpoint loader because the efficientcy loader is constantly changing its size and getting bigger and bigger. its a custom node group now.<ul><li>theres a vae input switch in this group: 1 = checkpoint vae, 2 = alternate vae.</li></ul></li><li>batch prompt switch on/off</li><li>Match Color group changed<ul><li>there are 4 ways now to use match color:<ul><li>use the colors of a loadad image via load image,</li><li>use a video file or</li><li>use images from a folder or</li><li>the input images (standard).</li><li>use one of theses methods, disable the others. if you disable all of them the input images are used. All of them using just one image because of the difference in size (batch_size).</li></ul><ul><li>the reason for that is, if you load previously created images from a sampler, match color is useless because the input images are rendered and not from the source material.</li></ul></li></ul></li></ul>Version 4<ul><li><s>almost perfect subject isolation by using a controlnet preprocesor.</s><ul><li>changed the masking area</li><li>added remove background</li></ul></li><li>adopted portions of the <a target="_blank" rel="ugc" href="https://civarchive.com/models/350524/jboogx-and-the-machine-learners-animatelcm-subject-and-background-isolation-via-invertmask-vid2vid-highresfix?modelVersionId=411119">v2 workflow</a> from jboogx<ul><li>added 2 ipadapters</li><li>added hires fix for the denoise ksampler and changed it back from standard to the efficient one.</li></ul></li><li>added a prompt scheduler. can be disabled by switching the conditioning input switch to 1. if you have to use multible queues to process all frames, batch prompt is not working or it starts from 0 everytime... the last time ive messed around with the start frame it did not worked out, but maybe its fixed. this could be something to try.</li><li>removed the group latents because empty latent was not working on the 1st Pass sampler like it was set up. instead, i have added individually adjustable vae encodes for both samplers (Denoise and 1st Pass). this allows you to chain them or disable denoise and start with 1st Pass. start with 2nd Pass would be dumb thats why theres none xd.</li><li>because an empty latent is something really boring I've added noise injection to it. ;)</li><li>both groups: remove background and controlnet replace background requires the group masking, so dont disable it.</li><li>check if skip: queue * load-cap is = 0 and set to fixed. sometimes i forget to reset this.</li></ul>Version 3.5<ul><li>added overlap (e.g load-cap = 100 overlap = 15 will render 115 images in all runs. skip images will always start at the load cap wich means that you have 15 frame on the end of the last run and 15 frames on the start of the current run that can be used for blending. on videos with 30 fps 15 frames will be 0.5 seconds for a cross fade).</li><li>load cap = 0 will render all frames/images disable overlap in that case (set value = 0).</li><li>fixed the evaluation of skip images: Skip: Fixed Value combined with the queue didnt worked so far. now you can use them together: queue + skip a fixed amount of images on start + overlap.</li><li>added a new aio controlnet group.</li><li><s>added conditioning input switch for the controlnets to keep it "clean", bypassed groups are not influencing the conditioning anymore.</s></li><li>added prompt everywhere in the prompt group in case all controlnets are bypassed, its inactive if 1 controlnet is active.</li><li>with low denoise values on the sampler you don't need controlnets (just a hint).</li><li>added video combine to get the upscaled result.</li></ul>Version 3<ul><li>rearanged some groups</li><li>added a lora stack loader</li><li>changed the wildcardencode to the wildcard processor node. (used <a target="_blank" rel="ugc" href="https://civarchive.com/models/20868/200-wildcards-nsfw-and-sfw">Wildcards</a>, extract them and copy to "\ComfyUI\custom_nodes\ComfyUI-Impact-Pack\wildcards", restart the server.)</li><li>changed the image blend node in the controlnet lineart/hed because at this point i ran out of mem a lot. the downside of this node is that it can not work with batches. thats why i converted the batch to a list at this point for the controlnet groups alone. it takes a while with a big batch. thats why i've added the softedge+ from a jerry davos workflow. i think its faster but i like my lineart/hed creation so i keep it xd. with both of them you can use the models: canny, lineart and softedge ;)</li><li>color match is now before the face swap node. face swap is doin a color match anyways (face only) but this allows to change the color with the filter adjustments and there is no rebatch needed in every group.</li><li>theres a new node for sharpen the image before applying face restore to sharpen the whole image. the images may no longer look too sharp on the face after the restore is done.</li><li>upscale is now the last node (skip this if you have a workflow for upscaling the video).</li><li>added noise injection to the latent upscale sampler group and the option to enable or disable both, the upscale model and the noise injection.</li><li>updated and tested a lot for memory usage</li><li>updated the preview areas and video combine and some other stuff.</li><li>the author of vhs (<a target="_blank" rel="ugc" href="https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite">video helper suite</a>) has changed the frame_rate of video combine from int to float. in old workflows you can fix this by changing the frame_rate from input to widget and back and replace the int node with a float.</li></ul>Version 2.1<ul><li>fixed a connection issue<ul><li>was running out of memory on the latent upscale group, loaded the created images from the ksampler via input image to rerun the whole thing (without ksampler (denoise)) and noticed that the latent upscale group didnt received an image input :( sorry for that</li></ul></li><li>connected ksampler (denoise) and 1st pass in case you want to run them all together.</li><li>changed the rebatch size from 1 to 10 (to much ram usage if its 1)</li></ul>Version 2<ul><li>scale image can be disabled now</li><li>added bookmarks for faster navigation (num keys 1-5)</li><li>added vae encode to use the original image as base and a ksampler with denoise only</li><li>added latent upscale on pixel space</li><li>changed the way rebatch is used</li><li>replaced video matting with a new node</li><li>changed the color match input for the reference image to one image instead of the complete input-batch</li></ul>Version 1<ul><li>supports wildcards and loras in the prompt, 2 controlnets, replace background (controlnet).</li><li>should work with low vram (using low rebatch values)</li><li>each group can be disabled</li><li>easy to use in steps.<ul><li>each group saves the results. so if you run out of memory you can load the last created images and continue where you left off. (simply switch off the previously used groups beforehand)</li><li>the overlapping technique sucks, so try to take as many pictures as possible in one go.</li></ul></li><li>Wildcards: <a target="_blank" rel="ugc" href="https://civarchive.com/search/models?sortBy=models_v7&amp;query=wildcards">https://civarchive.com/search/models?sortBy=models_v7&amp;query=wildcards</a></li><li>Motion Model: <a target="_blank" rel="ugc" href="https://civarchive.com/models/139237?modelVersionId=154097">https://civarchive.com/models/139237?modelVersionId=154097</a></li></ul>

AnimateDiff (Advanced), controlnets, face detailer, face swap, face restore, remove/replace background, ipadapter, vid2vid, img2img, img2vid

because theres still no fix ive removed searge node.its just a small fix.

Files with SHA256: d7bfc7c86cfe025112965afc9c1274cb59a86b50050168fe6d151214afb01c99