UPDATE: WAN 2.2 Smooth SVI Pro Workflow V.2 (03/25/2026)
Updated the workflow to use the ComfyUI-NAG. I tried to give more control over every video/extension so each could be costumized freely. You should have a lot of control now. Make sure to fully update your Comfyui and the extensions.
There are detailed notes now. Make sure to read them. :)
Download the videos on the Showcase and drag to ComfyUI for their full worklfow. I tried some different setups: 32 fps with no interpolation, 24 fps with interpolation, euler/simple and uni_pc/sgm_uniform... These are just examples on how you can mess and have fun with the setup. ^^
You can turn on and off Video Interpolation for each extension. Make sure your setup can handle Interpolation of longer videos.
I tried to make everything easy to find or change using very simple subgraphs. If you want to add or change something it should not be hard to do so.
Added a simple Node for saving the last frame of each video/extension. Some suggested this for helping with their video editing and creation.
The videos on the showcase used the following Model and Loras:
Models: Smooth Mix Wan 2.2 I2V v2
Light Loras:
HIGH - wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_1022.safetensors
LOW - lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors
SVI Pro Loras:
HIGH - SVI_v2_PRO_Wan2.2-I2V-A14B_HIGH_lora_rank_128_fp16.safetensors
LOW - SVI_v2_PRO_Wan2.2-I2V-A14B_LOW_lora_rank_128_fp16.safetensors
Used extensions:
Have fun!
======================================================
WAN 2.2 Smooth SVI Pro Workflow
Its a big workflow. Make sure to fully update your Comfyui and the extensions.
Used extensions:
==========================================================
Its divided in sections to make it easy to navigate.
MODEL SETUP
Here you select and set up your High and Low Models (safetensors or GGUF), your Light Loras, the SVI Pro Lora, your Shift for model Sampling, the Scheduler and Sigma and finally Clip, VAE and your prefered Sampler.
The videos on the showcase used the following Model and Loras:
Models: Smooth Mix Wan 2.2 I2V v2
Light Loras:
HIGH - wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_1022.safetensors
LOW - lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors
SVI Pro Loras:
HIGH - SVI_v2_PRO_Wan2.2-I2V-A14B_HIGH_lora_rank_128_fp16.safetensors
LOW - SVI_v2_PRO_Wan2.2-I2V-A14B_LOW_lora_rank_128_fp16.safetensors
There is a Note on the workflow with suggestion for values for Shift, Sheculer and Sigma. Be sure to check it!
VIDEO SETTINGS
Set your Starting Image and the video resolution. Its set to 720x480 as default.
VIDEO EXTENSIONS
Set how many times you want to extend your video. You can extend up to 5 times! Thats more than enough for 99% of people. More advance users will known how to go even further.
FIRST VIDEO/EXTENSION (1/2/3/4/5)
The 'NODES GUIDE' you help you easily find all you need to set up!
Set your High and Low Loras if you plan to use then
Set the total lenght - in seconds - for each video.
You can also set your Noise/Seed
Type your Positive and Negative Prompts!
=====================================================
VIDEO INTERPOLATION
I decided not to include video interpolation. Long videos makes it very common to users to run out of memory. If you really want to inerpolate video during the generations just add adn connect an interpolation node between 'Upscale Image By' and 'Video Combine'.
There are several free tools that can be used to do the same. The videos on the Showcase used Aiarty Video Enhancer (Paid), Flowframes (free) and DaVinci Resolve (free).
Just use your favorite tool for it! :)
=====================================================
Have fun!
Description
FAQ
Comments (38)
Just tired my first video using this WF . Omg amaizng. I just need to prompt a cumshot. As minee... well... looks like white paint lmao. A 10sec video took my like 5mins if that. Id post my first video just to embarrassed about tjhe cumshot lmfao 🤣
is the wf longer than 5s?
Same problem here :S
I fixed it by bypassing the set/get nodes for the sigmas and direct join from vaecodif to the initial video node
@redcloak90882 Hi, can you share your workflow? I can't get it to work.
@lolbleach001584 I'm sending it as soon as I finish, still making tests
Hi, the examples look very good, especially the skateboard scene!
Unfortunatelly the link to the low speed filter is broken, but I think I found it on huggingface.
Important note if you're coming from other SVI workflows to this one! Read if your results are turning out blown out and wrong!
Make sure you have the Kijai-optimized SVI loras! If you have the old VITA versions, they won't work right! They're linked in the workflow description.
I downloaded all the models and the workflow, but sorry, nothing works. It's not for lack of trying. When will there be a simple native model in ComfyUI without all this hassle?
It will appear exact the next day after you find out how it works. Trust me, I checked that earlier.
Thanks for your work!
I had to turn of sageattention to be able to run the workflow. I was trying to install it and its been a pain. Any guide or tutorial that you recommend to install it?? Thanks!
Thsi reddit link was what i followed. Check the comments as one guy talks about there some setup scripts he wrote but I didnt try that. https://www.reddit.com/r/comfyui/comments/1n8v3zy/detailed_stepbystep_full_comfyui_with_sage/
Maybe Im missing something but when i try to run a video with 3 extensions I just get 3 videos that all start from the same image
Same issue here. The 2nd video doesn't start where the first one left off. It basically just creates two videos without any continuity. Anyone know a fix? What could be wrong? I'm using the downloaded workflow.
I just did some more testing and it actually totally ignores the 2nd prompt. It just uses the first prompt and the starting image. Something is very wrong.
@lateralus in the second noise seed, change from random to fixed. The same for the next extension and so on.
@cookedleech212 Noise seed doesn't have anything to do with it. The issue is the extension nodes all have the sub graphs bypassed internally. It is literally hollow guts until you click edit and remove the bypass on all nodes inside. See the comment I left here.
I am pleasantly surprised by your work. True, I had to attach another model and lora to your workflow, but as a result, I got a good consistency of the character even at step 4. Thank you.
the rendered videos have nothing to do with my original input image. am I missing something?
sounds like you are using a t2v model instead of an i2v model
@futacrafts ah, yes it was by accident!
AMAZING job, great workflow (though i change it a little to my preferences). The best creator for noobs and novice. Finaly i can create 1 minute video without errors like forced loops or memory issue (Though i have a lot of memory).
This is my first comment in civitai, but i am so much happy and I clap my hands towards Digital Pastel for so much care for presenting his works for novice.
(Most creators have issue of having workflows with addons which is hard to download, but Digital Pastel include links which allow to download things manualy when automatic download have problems and for novice even 1 little problem can make workflows unable to work.)
great workflow, very easy to get running and well explained. 10/10
The workflow is very successful when producing a single video, but when producing multiple videos, it does not continue from the last frame of the previous video; it starts from the first reference image every time.
Same here. First I thought I did something wrong but I have the same results. Every Extra starts at the first image again. Changing Seed to fixed like suggested in another comment did nothing.
This has been doing great for me while I haven't been able to get other video workflows to work. Thank you!
If I were to suggest a version 2, incorporate last frame img2vid. Define a first and last frame for the initial generation, then let it start from the end of the previous video as it does now with the user defining the last frame of each extension.
That way we can basically create key frames every 5 second to direct a full coherent sequence.
Does anyone has an explanation video for img2vid? Not this workflow specifically but more of the technical side like lows and highs.
Hello! When it comes to ComfyUI in general ChatGPT and especially Copilot has a great hit rate for me. I'd say start asking questions there and then keep asking for clarifications as you go. I find it useful to start off with "Please explain this to me as if I was a complete beginner" to avoid drowning you in unfamiliar terms. :)
The recommended Lightx2v Loras in this workflow have been blurry messes that barely create any movement in I2V. Were they for T2V?
Edit: Ah, I think the High Lora requires a specific workflow and Sigmas variables to function. Though, if you suggest this particular Lora, it must have worked on yours?
This workflow is pretty awesome, only problem is the state it was saved in when it was uploaded. Other's get the "it keeps saving multiple 10 second videos with the same start". That is because by default the extension groups are bypassed, so you expect to only remove the group bypass but that doesn't work. Inside the custom Extension nodes the sub graph nodes are all completely bypassed. You have to "Edit" the subgraph for the Extension nodes, go in and remove the bypass on all nodes. No idea why it was left like that since all you have to do is bypass the group like they already were. This gets you progressively longer clips as expected.
Another issue with prompt adherence is that the author did not bring out the high and low CFG for the guidance scheduler, so they are basically hardcoded unless you Edit the initial/extended nodes. Just bringing the CFGs into the subgraph's input connector allows adjustment from the top level node. So that at least gives some weight adjustability to the prompt.
These staged SVP workflows are tricky so I need to mess with it more. With the above adjustments it's easily the best one I've found so far.
I thought that was odd when looking in the subgraph but I thought DigitalPastel did that deliberately given the examples above show prompt adherence. WTF? Thanks for bringing it up.
@DarkEngine2024 I don't know why it's like that. Bypassing the whole group without bypassing the nodes inside the custom node causes a nag error about "failed to validate" on the bypassed extension groups but it's benign. The model setup settings are also whacked out and it causes bad prompt adherence and jumpy behavior. Basically the LoRAs seem to get a super boost rather than what you expect from a strength of 1.0. It's a good workflow but the saved state that was uploaded is pretty rough.
Hmm. I was hoping this would fix the issue for me but when I checked the subgraphs they were all active. I'm still getting generations that just restart from prompt 1. What's weird is sometimes if I set up three prompts it will do the first prompt twice, then add in the third prompt on top of that. I wish they would upload a v2 with everything working.
whats the longest we can do for this ??
I tested this workflow a lot. I have great results with these settings : GGUF 8B - DR3EAMLAY V1 at 1.00 - Sampling 8 / 8 Sigma 8 / 4 2 extension activated 7s/7s/6s - I use an image 2560x1440 resize image in the workflow : 1592x896 (so fucking big). 3090rtx+32gi ram = 50 min. I get a 19sec video at 2560x1440. Quality is great. After that I use a workflow with GIMM VFI to smooth it at 48 fps.
certainly the best svi workflow I ve tested so far.
I have great results with these settings : GGUF 8B - DR3EAMLAY V1 at 1.00 - Sampling 8 / 8 Sigma 8 / 4 2 extension activated 7s/7s/6s - I use an image 2560x1440 resize image in the workflow : 1592x896 (so fucking big). 3090rtx+32gi ram = 50 min. I get a 19sec video at 2560x1440. Quality is great. After that I use a workflow with GIMM VFI to smooth it at 48 fps.
The author seems to have given up on all of us and doesn't comment on the problems.
Alright, next version is finally done! Thank you for all the feedback! ^^
Its a big update! Lots of changes but still follows the same formating.
After lots of experimenting with old and new Ksamplers I ended on the ComfyUI-NAG's KSampler for this Workflow. It showed to be the best in terms of good generation time and good quality. Its very reliable and gives great results. Make sure to install it, update ComfyUI and all extensions used on v.1!
ComfyUI have been updating a lot lately. I'm using ComfyUI v0.18.1 and everything is working fine!
Lots of control now: You can set the steps/sampler/scheduler for the inital video and all extentions separately. Do you want the first video to use 8 steps, Extension 1 to use 6 steps and Extension 2 to use 8 steps? Go ahead. It will work. I already tested it. I'll be sure to put videos made like this on the showcase. ^^
You will also be able to turn on and off interpolation for each extension freely.
Right now I'm working its UI to make it easy to read, modify and use - It will have helpful notes and links to the extension used.
Took a long time but testing SVI Pro takes quiet some time. Gotta make lots of long videos. lol
And I'll make sure the initial setup can work from the get go this time. >.<
I should be able to finish setting up everything and uploading tomorrow. ^^
Gotta finish this so I can move on to LTX 2.3! =D