This workflow focused on ease of use and powerful features for creating segmented videos.
Features:
Create unlimited videos beyond Wan 2.2 capabilities, thanks to SVI 2.0 pro. Easily add new sections to continue your video further, all in single continuous workflow
Automated save/loads. Continue your video from where you stopped, even after comfy restart. No quality degradation on video compression
Change base models for each section as well as LoRa. Find which ones work best for your idea
GGUF support for low VRAM systems. Use your GGUF clip and diffusion models like usual models at unified fields
Separable upscale/interpolation at the end. You can stitch a full video purely from saved data, preserving RAM and VRAM, or just generate from the beginning to upscaled video in one go
RAM optimizations. If your system can generate 1 segment, you should be able to do them all!
Chose different reference images for each video segment. It can introduce model to hidden assets/characters, that aren't visible on the start image. No LoRa required!
Precisely control every video segment: prompt, length, seed. Check for perfect stitch with previous generation only, no need to search through entire video
You can find guide for the workflow inside Options subgraph (v3.0), it's your starting point.
Feedback is appreciated! Share your thoughts
Version 3.0
See changes in version history
For this version you'll need my new custom nodes. If comfy manager can't recognize them automatically, you can find them manually by name:
ComfyUI-HeresyNodes
Or download them using git repository link
https://github.com/heresyCoder/ComfyUI-HeresyNodes
Remove old nodes ComfyUI-WanVideo-LatentIO used in V2 if they cause conflicts, they are added in the new node pack with some changes
April and Kit example was made using multiple input images. You can find them in the image gallery of v3.0
Note that workflow in this example was upscaled separately from generation, that's why all sections are bypassed.
If you want to use it, uncheck full_video_only parameter on options subgraph, unbypass sections and turn off upscale/interpolation/save on full video subgraph.
Or just download as JSON file
Version 2.0
See changes in version history
Based on kijai's SVI 2.0 workflow:
https://github.com/user-attachments/files/24359648/wan22_SVI_Pro_native_example_KJ.json
New custom nodes were made for saving latents of a video (Save Wan Video Latent) and loading from "output" folder (Load Wan Video Latent).
If you can't install them with comfy manager, you can get them manually:
https://github.com/heresyCoder/ComfyUI-WanVideo-LatentIO
Frieren with a mimic example was made using multiple input images. You can find them in the image gallery of v2.0
Description
Version 2.0
New features:
Refined design. Much less visual clutter
Now saving the last image of a section. For better navigation between outputs
Greatly reduced excessive video encoding for unchanged segments
Added fast action buttons for easier activation/deactivation of groups
Others:
Fixed flow order: previews now showing immediately after generation
FAQ
Comments (69)
This workflow is amazing
I am not able to get results anything like your example. I can generate the first segment fine but the second one ignores the first. The second section processes the loaded image like its own I2V resulting in the first and second videos having no transition. If I use the same image for both sections then there is a slight transition but its mostly a jump cut.
Did you downloaded SVI 2.0 pro LoRa? It goes to global LoRa on control panel
https://huggingface.co/Kijai/WanVideo_comfy/tree/main/LoRAs/Stable-Video-Infinity/v2.0
It probably requires weight adjustment based on base model (with lightx or without), but I keep it at 1.
I'm also getting issues with transitions from time to time, it's not perfect. Adjusting the prompt helps in many cases of inconsistencies
I'm using that lora. I've tested this some more by adjusting loras, images, and weights. No matter I do the second section is completely separate from the first. There is no continuation from section 1 to 2, it just does a slightly blurry jump cut between the two
Could you share the prompt for the first generation? I can't get the character into the frame. Either the model draws a character based on my prompt with details, ignoring the first anchor_samples, or (if I don't describe it in any way, but only write "character") generates it without it at all.
I usually prompt something like this:
A girl comes from left off-screen and ...
Short description of a character after that would also help.
Check that your prompt is feasible within the context of your scenery.
It's not working 100% of the time, I'm getting slowly appearing ghosts pretty often.
Just keep in mind that images placed in anchor_samples are not something mandatory or strict, as end frames. They're just helpers, while proper prompting makes most of the work
You should make a tutorial because i tried and tried to make it work but it's impossible.. every generation is messy, it generates only the firts frame if i'm lucky.
You can start with my example. Download, drag and drop the video in your comfy to get exact workflow used. Download the images that I've posted here and place them according to workflow screenshot.
See how it works, maybe you need to adjust something
If there is big demand, I could do a detailed tutorial
Forgot to mention. I've encountered two bugs related to preview videos. If you're generating from newly started comfyui, then previews may not show at all. Also generations can randomly give you only the first frame. Both bugs are on GUI side and can't be fixed via workflow. You can switch between workflow tabs to workaround this. Videos must play after that in both cases.
Maybe you came across this bugs too
Hi, could you please implement the perfect resolution find node? I'm a noob and I don't know how to do it.
Hi!
You want to enter the biggest side in pixels and calculate the other side to keep aspect ratio?
@vamorand yes
@lolbleach001584, you're right, that would be more convenient. I'll make it in the next version
@vamorand thanks
Due to capacity issues, I used the "Wan2_2-I2V-A14B_fp8_e4m3fn_scaled_KJ" model, and ran it with or without applying the default LoRa, but the video is generating a lot of noise. Is there a specific model that can be used?
lightx lora weights may be the case.
My examples are for Smooth Mix v2.0
Try to set strength of lightx loras to 1 high, 1 low.
I'm using this configs, depending on what I want to achieve:
Smooth Mix v1, no lightx (baked) - fast NSFW action, bad SFW
https://civitai.com/models/1995784?modelVersionId=2260110
https://civitai.com/models/1995784?modelVersionId=2259006
Smooth Mix v2.0, lightx with 3 high, 1.5 low - semi-realistic and drawn charachers, average at everything
https://civitai.com/models/1995784?modelVersionId=2513182
https://civitai.com/models/1995784?modelVersionId=2513186
WAN 2.2 Enhanced NSFW | SVI | camera prompt adherence, lightx with 1 high, 1 low - best adherence to reference image, realistic, good male anatomy, slow motion
https://civitai.com/models/2053259?modelVersionId=2609148
https://civitai.com/models/2053259?modelVersionId=2609141
Thank you for this excellent workflow, I've been looking for SVI pro workflows all over the place and all of them seem extremely complicated, This one is simple enough that even a dummy like me can get it to work.
I'm glad you like it 😊
Next version should be even more straightforward and powerful, with some unique features among SVI workflows. But it may take a while to complete
I love the long version of the workflow, thank you. By the way, maybe I'm using it wrong, but if I create a 5-second x 4-chapter video, and I create Chapter 1 on the first day and Chapter 2 the next day, if I load RAW from Chapter 1 on the first day and then enter the prompt for Chapter 2 to generate it, it ends up being a 15-second video. If I turn off Chapter 1, it won't generate. Imagine if I wanted to create a continuation of Chapter 1 on the first day, and then enter the prompt for Chapter 2 the next day, it would continue from Chapter 1. Is it possible to switch Load RAW to start from any chapter, and then generate the continuation from Chapter 3 or Chapter 4?
If I understood you correctly, that's exactly what I'm doing in a new version.
To load data and start from section 1 every time is not very convenient, because it requires to move your prompt to section 1 from wherever it ended up previously, so being able to save/load from any section is my goal for the next version
@vamorand Great, I'll wait.
Hello, thanks for sharing, it looks so cool!
but i took your workflow with the video for example but i only have a character walking on a white background, not the scenery. I must have forgotten something... the base image in the group node "new generation" seems to be ignored ...
Hi!
I don't get what can cause this. Rechecked by downloading my uploaded images and workflow in video. All works perfectly. I have sage attention 2.2 installed, booting comfy without it works too, if base model loader node not activates it by itself at least.
Try similar workflow, can you get it to work?
https://civitai.com/models/2325986/wan-22-svi-pro-20-i2v-for-12gb-vram-multianchor
nope, disabling sage attention (for real now) doesn't do anything meaningful, unless slowing generation 2x times
Yes, it's Sage Attention, it seems to be causing an error; I need to reinstall it. Thanks for the help <3
Have mercy on a poor fool but I just can't get this to work, I've installed the custom nodes with the exception of the Save/LoadWanVideoLatent as they're not showing up in search or install custom nodes in the manager but since they're not in the new video workflow I hoped it was ok that they're missing?
Secondly I'm getting the Sage Attention missing error but changed from auto to disabled as I understand it's just for speeding up generations but should still work when disabled?
Finally the Diffusion Model Loader KJ high/low nodes are missing the Wan22_smoothMix_V20_HIGH/LOW.safetensors files and I can't find a checkpoint with the same filename but tried using Smooth Mix Wan 2.2 14B (I2V/T2V) High/Low checkpoints but while I'm not getting errors when I run the workflow any more I'm just getting noise in the output, looks like someone boiling melted marshmallow. I realise this is all probably beyond me but if anyone can give me some tips for getting this working I'd be eternally grateful.
Hi!
Save/LoadWanVideoLatent are still there, but can be omitted, if you don't need to save and load for continuous videos. I'll search what's wrong with them, as I need my custom nodes in the workflow
Workflow works without sage attention, that's true. But I highly recommend to install it. I've got like 2 times the difference the last time I tried to turn it off. Version 2.2 btw. It can affect quality by very little, but I can't tell the difference
As for base models, you got it correctly.
By default this workflow uses smoothMix_V20 with lightx2v lora (strength 3 on HIGH, 1.5 on LOW), that allows you to set number of total steps to 4 (I'm using 6 for slightly better quality). If you're not using this lora, you should set at least 20 steps, separated by 10 between LOW/HIGH (can be found on control panel at the right side)
You can get lightx2v lora here, if you're not using it already:
https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors
Save/LoadWanVideoLatent nodes can be installed manually using this link from workflow description:
https://github.com/heresyCoder/ComfyUI-WanVideo-LatentIO
@vamorand thanks so much for your help I really appreciate it!
I'll try more steps and hopefully that makes the other lora work. I appreciate the tip about sage attention, I will try to get that working as well since it sounds like a huge speed boost. I can see where you set the steps on the right side of control panel but I'm not 100% certain what you mean by 10 steps between low and high, do you mean putting a value of 10 in the SplitSigmas node below the BasicScheduler node? I'm sorry to pester you with my problems I'm extremely grateful for any further assistance you're able to give!
I tried to manually install the latent save/load from the Git link but I'm still not quite sure what my mistake is. I followed the ComfyUI guide to manually install a custom node and used cmd in the custom_nodes directory to git clone the ComfyUI-WanVideo-LatentIO link then followed the next step in the ComfyUI guide for manual custom node install to use
python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-WanVideo-LatentIO\requirements.txt
to install the dependency but I get an error in the terminal since requirements.txt doesn't exist in the ComfyUI-WanVideo-LatentIO folder so I'm still a little stuck on how to get ComfyUI to find those custom nodes and stop giving me the message that they're missing when I refresh even though the folder is in the custom_nodes directory.
Again thanks for trying to help me, your workflow is exactly what I was looking for, a way to apply different loras at each stage of a wan22 svi2 generation so that I can apply a new lora every 2-3 seconds in a long video.
@mistergoodbytes464 ,
>do you mean putting a value of 10 in the SplitSigmas node below the BasicScheduler node
yes
>I get an error in the terminal since requirements.txt doesn't exist
Posting nodes is new for me, so errors can occur. I'll check it later, thanks.
I'm trying to make a simple workflow without typical comfy's cluttery for everyone to use, so it's ok to point what causing issues, so I can address them
@vamorand your work and assistance is appreciated!
Is it normal that if I don't reuse the character's image in clip 2, the face loses consistency?
You have 2 sources of consistency:
1. Last frame of previous section
2. Reference image (anchor sample)
If character's face is visible when section starts, it will be consistent through that section. If it's not fully visible, then you must put it in reference for that section to save consistency. You can put many assets into reference image, make a collage in some photo editor. I believe we'll get multireference someday as paid video models now have, but for now we can workaround it using collages
A very good workflow!But i meet a problem when i try to use the saved raw data. My raw data is in C:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\output\Video\2026-02-24\193551,when i hit "run",an error occured,it said"directory is not valid: G:\Programs\Comfy\ComfyUI\output\Video\2026-02-24\193551/frames". What should i do to?
Set your comfy path in options subgraph instead of G:\...
I'll obtain this path automatically in the next version, there are nodes for that
@vamorand Thank you! Problem solved!! :D
Works really well! Makes me wish we could add even more stuff to the "context collage" since it's so powerful. Couldn't figure out the "Load raw" option due to the folder paths not matching mine (couldn't find all the nodes for this)
This is exciting to see. I've been slightly editing a workflow for a while now, but I always wondered if it was possible to add multiple image references for a generation. Which is what you do I'm not mistaken. Can't wait to try this out and see how you set it up. I do a lot of long videos so seeing new things is exciting!
It's one ref per sampling process. And multiple if you count whole stitched video. Doing many references per single sampling would be a gamechanger. Interface can accept them, but doing nothing with refs more than 1. I even tried to vibecode my custom SVI nodes, but with no results. This is technically impossible yet, unfortunately. Maybe next SVI version could do it, or entirely new base model
Dang. So is editing photos onto one kinda the cheat code for now? Saw that you placed part of her body and the mimic on one reference.
@PopcornMaven, yep. at least refs don't have to be perfect as last image approach. just put some objects there, model can figure out how to place them according to prompt and overall context
Comfy cant find these 3 nodes despite trying to manually install, use manager, or git pull them?? any idea why?
JWStringConcat
JWIntegerSub
JWIntegerAdd
Uhm
I don't have nodes with this names in my workflows. Closest I can find is jamesWalker55 nodes (String Concatenate, for example). No idea what's wrong
GitHub - jamesWalker55/comfyui-various
i tried the exact workflow from the video file, no parameters modified, with the wooden interior as base image, and the elf girl image as anchor image of sections. but the output looks twisted and blurred, the background seems like a snowy ground and an elf girl is blurring in misery. i tried offical wan2.2 checkpoint and smoothmix checkpoint, both give twisted output.
i also tried KJ workflow, it gives output correctly.
is there anything possibly wrong with my local setup with your workflow?
Lightx lora and steps (sigmas) must be set according to base model's recommendations. Also sage attention can be the case. I'm using sage attention 2.2. If you don't have it installed, you may need to install it or disable it in diffusion model loaders nodes
@vamorand thanks, i'll take a look.
Yeah, on my short time using Wan2.2, I've realized that blurry/snowy videos usually come from mixing a checkpoint that already has a lighting lora baked in and using a lighting lora on top of that... or, just using a checkpoint with no lighting lora with very small steps.
post a you tube tutorial for those of us who are slow please
Great workflow. Is there any way possible to not downscale the images? For example in the standard WanFirstLastFrameToVideo node, your start image input and your sampling width and depth inputs are all separate. So you can use a 1440p starting image to make a 720p video, and this gives it a very good quality boost.
But with all of these SVI workflows the sampling resolution is based off of the image resolution. Anyone would then need to downscale a 1440p image to 720p otherwise you'll get OOM, but then you lose lots of quality since your starting image becomes 720p. Surely there must be a way?
most of svi workflows based of KJ's workflow, which is using SamplerCustomAdvanced node for sampling. it doesn't have separate width/height. do you have an example workflow, that uses this technique to increase quality?
@vamorand No I've not yet seen an SVI workflow which can do that. I'm starting to think the SVI node would need that functionality for it to be possible.
Separate question for you -- if using your default settings as per the workflow, and the motion_latent_count from the SVI node is set to 2 -- what is the correct amount of overlap and 'previous length' to use in your 'settings' subgraph? 9 and 72 are very close but there is a very slight but noticeable blur during the transition.
@m55971 ,
>I'm starting to think the SVI node would need that functionality for it to be possible
well, if it is some vae encoding tricks, that still may be possible, because sampler requires encoded latents. that's why I want to know about workflows, that can do this.
>if using your default settings as per the workflow, and the motion_latent_count from the SVI node is set to 2
I had no luck with options other than 1, so I didn't tested it and can't help. maybe I did it the wrong way back then.
>what is the correct amount of overlap and 'previous length'
overlap is a default setting, I didn't experimented with it either. previous length used solely for preview of the first section of loaded video, generation not affected by this
@vamorand
Got it, thanks.
>well, if it is some vae encoding tricks, that still may be possible, because sampler requires encoded latents. that's why I want to know about workflows, that can do this.
The basic ComfyUI Wan 2.2 FLF template can do this:
Thank you for the V2.0 workflow. I am trying to use the save/load raw option. Say I generate Section 1 and Section 2 from scratch using save Raw option. Now if I disable section 1 and section 2, enable section 3 and click the toggle to Load Raw, I get an error saying
* ImageBatchExtendWithOverlap 1052:859:
- Required input is missing: source_images
I have the correct latent folder path setup. Please let me know if I am missing anything.
It's not gonna work like that. After loading data your section 1 is now continuation of section 2 that you saved previously. Section 1 must be always active, no gaps.
I'm making workflow that handles save/load by itself, where you just continue long chain of sections. but it's surprisingly complicated, so I need more time to make it
Thank you for your work. Im trying to use this workflow for comfyui Stability Matrix but I have to modify a few nodes not compatible with this version (SM Comfyui uses Python 3.12.x). Copilot is my guide but still not sure if that means that Im going to change the workflow into a Frankenstein node work. Do you have any recommendation to substitute comfyui-various nodes?
I was able to add individual nodes from comfyui-various and was able to use the workflow as is. The problem now is that I can't use triton/sage for acceleration so it takes almost 6 minutes per video and I'm only getting white noise that looks like water. I'm checking which parameter I should change to fix that.
@RokkumanRaito, disable sage attention option in the base model loaders nodes (Diffusion Model Loader KJ). but that's slow without it, yeah
@vamorand Thanks for the answer. After checking some parameters I was able to create some videos.
Thank you for the great workflow. I learned a lot from your workflow and helpful instructions. This method seems to be one of the most versatile tricks for video creation using WAN 2.2 I2V.
After the first five seconds of the video, it gets brighter every five seconds. Can someone please help?
I never had this issue before. It can shift colors slightly on transitions, but still ties to previous segment and reference images. maybe parameters for your model setup are incorrect (cfg, step count)
@vamorand I agree, I must be missing something important. Could you maybe give me some hints on what I should change? Also whenever I use lightx2v as a low LoRA I get incredibly distorted results. Sorry for being such a newbie but ChatGPT wasnt much help
@vamorand So apparently, I was using the wrong version of SmoothMix. Now everything is fine. bruh
Could you put a link to the wan22 smoothMix
Unless you renamed it, I cannot find it
@vamorand Thank You!!!
Going to play with this tonight. So there is no prompting?
metadata is in video. I discovered that drag and drop is broken in my current comfyui setup. maybe that is the reason you can't find it. open file from menu works
https://www.reddit.com/r/comfyui/comments/1tjiwuc/drag_and_drop_doesnt_works_anymore_in_recent/