SVI Extend
https://github.com/vita-epfl/Stable-Video-Infinity/tree/svi_wan22
Create videos and extend them seemlessly using SVI.
Following SVI LoRAs are mandatory:
switch between default behaviour, anchor_samples and end_frames within the same subgraphs
connect an image to a part and enable the respective toggles to use end_frames or anchor_samples
NEW! v3
Extend existing videos using https://github.com/wallen0322/ComfyUI-Wan22FMLF
enable "video extension" toggle inside the settings
uses source video resolution by default
rescale video using the megapixel slider by enabling "video rescale" toggle
use included version of the nodes from inside .zip or download the latest version straight from the git if issues arise
More info inside the workflow.
AIO i2v+t2v
All in One workflow for for basic WAN 2.2 video generation.
Following features included:
Switch seemlessly between 2 and 3 sampler solutions
Toggle between i2v or t2v
Postprod
Facedetailer
uses t2v Model + LoRA for inpainting - resources needed included in workflow
Toggle between GIMM VFI and RIFE VFI Interpolation
Upscale
Tensorrt Upscale with Model
Basic Video Upscale with Model
RTX Video Super Resolution Upscale (insanely fast for decent quality)
Frame Clipper
Seamless Loops using custom RIFE nodes https://github.com/Artificial-Sweetener/comfyui-WhiteRabbit
Upscale + Interpolate
I recommend using this workflow instead of upscaling with the generating workflows, since you never really know what kind of results you get, ending up upscaling a bad video and wasting time. I included toggles so you can't use multiple interpolation or upscale nodes at once by mistake.
This includes:
WAN Facedetailer
use any WAN 2.2 T2V low model + the following T2V LoRA:
lower resolution from 768 to 512 if you have VRAM issues
Put the following file into "ComfyUI\models\ultralytics\bbox":
WAN Refiner (massive VRAM cost)
increase denoise if you want more inpainting
Sharpen, Gamma, Brightness and Contrast control
Frame clipper (remove unwanted frames at the start and/or end)
GIMM VFI + RIFE VFI interpolation (I recommend GIMM VFI, much higher quality but also much slower)
Tensorrt Upscale + Basic Video Upscale
both use basic image upscaling models
Tensorrt (faster than Basic Video Upscale) with AnimeSharp4x is recommended for anime
RTX Video Super Resolution Upscale
insanely fast
decent quality
FlashSVR + SeedVR2
experimental
Video upscale models that are more intricate than basic image upscaling models
haven't had great results for anime yet
takes a LOT longer
Saving last frame for manual extensions
mmaudio
added Audio combine node
combine audio from an existing video with the generated audio on top
generate nsfw audio with the nsfw model and then combine that video with another generated audio track from the base model for background noises
removed interpolation for easier and faster audio generation - you have the following options:
upload raw unupscaled video to MMAudio Video node and upscaled video to Combine video node
upload upscaled video to both nodes but lower custom_width and custom_height of the MMAudio video node to about half for faster generation and to prevent VRAM issues
upload raw video to both nodes and upscale afterwards
Inspired by https://civarchive.com/models/2137833
Following resources necessary (ComfyUI\models\mmaudio):
https://huggingface.co/Kijai/MMAudio_safetensors/resolve/main/mmaudio_synchformer_fp16.safetensors
Description
Archive of the deleted workflows if someone wants them for whatever reason. I don't recommend using them anymore since I won't update or support them.
FAQ
Comments (26)
Is it possible to add "Last Frame" for each Extended video/First Video? This workflow so awesome but cant make it nicely loop
No, that's unfortunately not possible with SVI extensions. If you want to do loops then you need to use WanFirsttoLastFrame like in my other i2v workflow and do that manually.
It seems like someone made a SVI Node that actually allows to add end-samples https://github.com/Well-Made/ComfyUI-Wan-SVI2Pro-FLF
I'll play around with it and see if it's worth adding, I'll update the workflow soon if that is the case.
@LastAssignment Whoa man, thanks a lot! Will check it out too
Hey, I'm getting the error "SyntaxError: Unexpected non-whitespace character after JSON at position 4 (line 1 column 5)." I thought it was because I was having problems installing Tensorrt, but I believe I figured out how to install it properly. Any idea what might be the problem?
That error doesn't tell me enough to help you troubleshooting :(, do you have any node that pops up red when executing the workflow and if yes which is it? If it's tensorrt then try the following:
Enter "ComfyUI_windows_portable\python_embeded" folder run cmd.
python.exe -s -m pip install wheel-stub
python.exe -m pip install --upgrade pip setuptools wheel
python.exe -m pip install tensorrt
Do you have any suggestions for addressing lip deformation, specifically transforming from anime-style to a very real-life, fiery red lip?
The only way to effectively do that is to keep LoRA strengths as low as possible and keep the amount of extensions low if you're doing blowjob scenes. Most LoRAs aren't trained on anime so transformation is inevitable if the strength is too high, strengths may need to vary depending on the seed, scene, style etc. so it really needs a lot of playing around to get just right.
To add to this, you can also try adding this LoRA https://civitai.com/models/2222779/anime-style-wan-22-i2v at ~0.3-0.4 strength on low and high and don't forget the trigger "An1meStyl3, AnimeStyle". This may lower motion but it could be worth the trade off depending on the scene.
@LastAssignment Thank you for your suggestion. I'll give it a try.
Hi bro! How do you think is it better to use finetuned model with baked loras kire Wan 2.2 Remix or basic Wan 2.2 and your own Loras gives better results in general?
Hey, depends on what your goal is. Base model will definitely have the higher potential for quality but it will need a lot more fintuning with LoRA strengths and other settings to get just right. If you're willing to experiment with many MANY gens then I definitely recommend going for the base model and do the stuff yourself but if you just want some quick clips then you should just use a mixed checkpoint.
@LastAssignment I make videos with fictional characters, usually around 3 minutes long. Basically, they’re compilations of 5-second clips, about 30–40 clips per animation. Sometimes I do 10-second clips by taking the last frame of the first clip and generating a continuation, but obviously it almost always comes out a bit rough.
That said, I still get a lot of views on adult sites. You can check out my videos on xHamster if you’re curious — my username there is the same as here.
I’m using WAN 2.2 Remix and Fast Move with a baked LoRA, but I still add another LoRA on top. What I really want, though, is to extend clip length to 10–15 seconds without losing quality. I was looking for a suitable workflow, but honestly I’m too dumb to build something like that myself lol — and then I found yours.
I’m going to test it today and tomorrow. It looks very promising: really well put together, clear, and I really like the example clips you render. I hope this is exactly what I’ve been looking for. If everything works the way I expect, I’ll definitely thank you properly with Buzz!
I also tried SVI, but I didn’t like the results — too little movement. To be fair, I didn’t spend much time with SVI, so it’s very possible I was doing something wrong.
Thanks for sharing your workflows — you’re awesome!
@LastAssignment also i saw some Painter node, it's like a sampler i believe and it can understand 10-15 frames from previous clip, maybe it can give even better results if you'll put it into your workflow? But i'm not to smart as you in this, it's just a theory😅
@Honeyphoria SVI can be great but WAN 2.2 mostly held back by the lightning LoRAs hampering motion and reducing prompt adherancy. Base models using 3 samplers generally have the most potential but getting that ratio just perfect is really damn HARD without having a lot of blur left.
SVI really expects to have a good transition between extensions, this means you may need to regenerate individual parts by changing the seeds and see if the results are better. This workflow generally makes that stuff relatively easy if you have enough memory to keep everything in RAM at once (>96GB for longer videos, about 30s or 5 extensions, with FP8 model), you can still generate longer videos but once you hit that cap you'll need to regenerate the whole video again if you aren't satisfied which ends up being a huge timewaste.
I'm honestly still learning and experimenting myself but there's no 100% perfect solution atm :(
I've been playing around with painteri2v a lot but it doesn't work for SVI extensions and it generally adds a lot of unwanted stuff that I couldn't fix.
@LastAssignment Interesting👀 Thank you for the detailed reply!
This is excellent, thank you so much! :)
Bro, i need your help so bad! Here is a screenshot with my setting and models. For some reason I’m getting gray videos as the output. I tried rendering with the SVI model as shown in the screenshot, and also with the fine-tuned WAN 2.2 Remix with baked-in LightX — of course after disabling the LightX LoRA nodes beforehand. In all cases, the output is gray videos.
I’d be very grateful for your help, my friend! 🙏
You need to set base steps to 0 instead of -1
@LastAssignment I’m such an idiot — I read the note as not above 0 instead of below 0 😁
Now everything works, and it works perfectly!!! No harsh seams between the 5-second segments, perfect color matching — this is exactly what I’ve been searching for for so long. You’re an absolute genius!!
Words can’t express how grateful I am. I sent you 1000 Buzz — that’s all I had on my balance, but you honestly deserve way more!!
Can I ask one last question?
Is there a way to add preview renders to this part of the workflow — for each segment and also for the final output from the full merge node — so I don’t have to manually check them in the file system? I’m working on a server and that’s unfortunately not very convenient
@Honeyphoria No problem :) Thanks a lot! You can enable the save individual parts toggle, go into the subgraphs (https://imgur.com/a/XjJmbi0 First Video and Extend Video) and right click the "Output" nodes and click on "Show Preview", same for the "Full Merge" node at the end. This will enable previews for all parts and for the full merge at the end.
The examples look interesting. It's the first workflow I see with some "facedetailer" attached.
Keep up the good work!
The only workflow that can generate good eyes especially in loops. I tried many, really MANY WAN comfyui workflows in the last months and this one beats the rest at the quality.
The only thing you could change is to add an option to upscale without tensorrt in the AIO workflow, cuz that tensorrt just throws some errors, but in rest all good.
Congratulations!
I really began to think that the latest ComfyUI updates broke the WAN 2.2 quality, because I don't know why the workflows that were good til recently, they all stopped generating good stuff all of a sudden.
And smart move to not add that "Color Match" node, gj!
@TekeshiX I'll rework it soon enough, I was mostly focusing on the SVI workflow atm. Color Match is a curse, haha. It's really not made for video, it just fucks with everything instead of actually help. If you wanna fix tensorrt then you can try the following:
Enter "ComfyUI_windows_portable\python_embeded" folder run cmd.
python.exe -s -m pip install wheel-stub
python.exe -m pip install --upgrade pip setuptools wheel
python.exe -m pip install tensorrt
@LastAssignment One curiosity. In the WAN 2.2 AIO v1 archive there are 2 workflows:
1. WAN 2.2 AIO v1 WF - example video.json
2. WAN 2.2 AIO v1 WF.json
Is there any difference to them besides the "date modifed"?
@TekeshiX There are some small insignificant fixes to the normal workflow but they won't change the functionality, if it works then they are identical. I only added the example video WF for the settings used