This is a clean and concise workflow for ComfyUI that allows you to generate longer videos by chaining together up to 4 separate clips.
V3.0 now includes SVI for better consistency in longer videos!
Note: Full instructions, advice, and links to models can be found in the note in the workflow itself.
This workflow is based on my clean 2+2[+2] lightning workflow that can be found here:
You can use this workflow to generate a single WAN 2.2 video or you can have it chain up to 3 additional videos based on the final frame of the previous clip. In this way you can create longer videos, with distinct changes in action or framing without needing to generate the whole thing in one go. Simply fine tune the first clip with the others disabled until you are happy. Then don't change that clip and move on to clip 2. ComfyUI won't spend time regenerating the earlier clips each time you want to tune the next one!
Please consider tagging this as a resource you used if you generate anything with it and upload it here. I'd love to see what you're making with the tool!
Description
FAQ
Comments (21)
For some reason it won't do the last clip. I have it enabled, but it always stops after the first 3
Did you choose a file name for the save file? Defining that path is in the instructions.
The other possibility that I came across is if your clips are too long and upscaling is on then I often run out of memory when combining the final clip and it fails.
I think the latest ComfyUI 0.4.0 broke Anything Everywhere. I can't get the first step to work without manually connecting nodes.
Ah that's possible - I haven't updated to 0.4.0 yet so I might wait :)
I don't think the problem is with the ComfyUI. I have 0.3.7, same problem, must connect nodes to get it to work.
@TruePepega Fair enough! There could be a mistake in the .json when I exported it, although I think it might be to do with the Anything Everywhere nodes themselves. I don't think they like being imported very much and I should probably avoid using them - but they make everything so clean!
I think Nodes 2.0 update kinda broke it a bit.
@TheFatController I found the issue. I changed the UI language of the ComfyUI to English, and the workflow worked correctly. I assume the problem is Anything Everywhere does not recognise foreign words, thus can't connect the nodes. So changing the UI to English worked for me. Thank you so much for this workflow!
@TruePepega Ahhh, that's very useful to know! Thanks for commenting :) Glad you like it! I'm really interested to see what people make using it so if you upload anything please tag the workflow as a "Resource" so it shows up here and I can see it :)
@TruePepega Well dang! That worked. My language was already set to English, so I set it to something else, and then back to English. It worked.
Man this has been this best workflow i've come across for long video that isn't complete spaghetti. I used the wan 2.2 stable video infinity loras and it seems to really help to reduce color degradation and face consistency. However the workflow is good on it's own.
Is it possible to get an updated longer workflow? I tried doing it myself, but i dont know what tool your using to connect the nodes, i cant see the connections. But an extended flow so you can work into 8 clips instead of 4 would be dope.
There are "Anything Everywhere" nodes hiding under a lot of the nodes to connect things, that might be what you're missing!
I might take a look at the stable video loras, thanks for the tip.
I've also had more luck taking the hit on longer generation times and using 2 KSamplers (just bypassing the middle lora loader and middle KSampler in each step) but then turning off the lightning loras and going for 12+12 steps instead of 3+3+3 with lightning. The end results are a lot clearer when chaining the clips!
@TheFatController Between the step nodes, and how the anything everywhere node works I'm lost. But I'll keep at it,and I'll take your advice. Thanks Alot.
@TheFatController How do the step nodes under the sampler work? Is that just a modelsampling node?
@Superelmostar675 Hey! Sorry we might be on different timezones :) Which ones do you mean exactly? The "steps" number in each KSampler node? If that's what you mean, they should all have the same value for "steps" within each group. This is the total steps for generating each short clip. Then you want to have the "start_at_step" and "end_at_step" for each KSampler to line up. E.g. if you have 9 total steps, and you want to do 3 steps for each KSampler, you put "steps" as 9 for all three KSamplers. Then you'd put "start_at_step" as 0 and "end_at_step" as 3 for the first KSampler, start at 3 and end at 6 for the second sampler, and start at 6 and end at 9 (or 10, or 10000, just something bigger than 9!) for the last sampler.
Does that make sense?
I've been using 3+3+3 = 9 total, with the first one being high noise and no lightning lora, the second being high noise with the lightning lora, and the third one being low noise with the lightning lora.
And yes - the "Anything Everywhere" nodes can be confusing. I still screw them up sometimes when making complicated workflows that might want certain groups turning on and off.
A good guide is to disable Clip 2, Clip 3, and Clip 4, and disable the upscale groups until it's working. Then enable Clip 2 and try again. If it works and you like the output all the way through to however many clips you want then I usually turn the upscaling back on and run it again (upscaling can add a lot of time and you only want to do it once you like the results!). If you do that, make sure you set the seed node to "fixed" or it'll re-generate everything each time.
@TheFatController Thanks I got it all working. It took sometime. Thanks a ton!
It works well, thanks,
- one question : As we generate the second or third videos, it uses the last frame of the previous video right? so.. the current video starts with that last frame of the previous video.. then doesn't it mean we have a duplicate?
In summary, Previous_Last_Frame = Current_First_Frame, then we have two identical frames? In this case, I think it is better to remove one of the frames to make a smoother connection between the videos.
I believe you are correct, however when interpolating up to 60 FPS I haven't really noticed the duplicate frame at all. Additionally, I can't check right now, but I might have actually got it to generate from the final frame, but then skip that out during the combination. You'll have to look at the image batch nodes to check!
I did a slight change to use the GGUF models. That works fine for the single step workflow.
But here in the long extended version it simply crashed ComfyUI on the second step. No error message from Comfy, just an "Exit 0". I even tried smaller GGUFs but this doesn't change anything.
I do have 16GB of VRAM.
Anybody experiences the same?
Are you using the lightning loras for the 2nd and 3rd steps? You could re-download them in case one is causing an issue perhaps? Or just do a test with it disabled.
Does it crash without the GGUF models? I also have 16 GB of VRAM and I don't use them.
@TheFatController I also did get some crashing with variations of WAN 22 GGUF in Comfy, but I think I tied them down to higher bitrate GGUF pushes (that I can run in standalone) mixed with over pushing the Latent length, causing overflow errors.
Save early, save often, but eventually taking things down a bit and not trying to push so hard helped a ton.
I also ended up getting more overall stability and cleaner output out of using the WAN 22 14B FP8 scaled core models, GGUF was stable and working after some fixes but the huggingface model I was using was altering prompts a lot more because I lost some of the basic input items I was relying on on the Civit 14B safetensor