Workflow Overview: Lightspeed Wan Series
These workflows are optimized for the Lightspeed Wan 2.1 and 2.2 Image-to-Video (I2V) models by darksidewalker. They are categorized into the streamlined SVI-V2-Pro series and the feature-rich Standard (v12 and below) series.
1. SVI-V2-Pro-I2V (Barebone Series)
The SVI series is designed for users who want a clean workspace without unnecessary overhead. It focuses on the core I2V logic with two critical utility modules.
A. Resolution Management (Resolution Master)
To ensure compatibility with Wan 2.2’s specific latent requirements, the source image must be precisely resized.
Step 1: Load the source image into the Load Image node.
Step 2: Use the dropdown (next to the TV icon) to select a target resolution (e.g., 480p, 720p).
Step 3: Click the TV Icon to activate the selection.
Step 4: Under 'Actions', set the slider to 16 and click the Snap button.
Note: This "snapping" ensures the dimensions are multiples of 16, preventing tiling artifacts or model errors.
B. Temporal Smoothing (Interpolation)
Located in a dedicated group, this module handles frame rate conversion.
Primary Use: Converting low-FPS raw generations (e.g., 16 FPS) into smooth, high-fidelity output (e.g., 60 FPS).
2. Standard Wan Workflows (v12 and Below)
The legacy/standard workflows are "all-in-one" solutions, containing a broader suite of post-processing and utility tools:
Image to Video
The core generative engine for Wan 2.1/2.2.
FLF Loop
Facilitates seamless looping by utilizing First-Last Frame logic.
Downscale Image
Safely shrinks oversized source images to prevent VRAM OOM (Out of Memory) errors.
Extract Last Frame
Pulls the final frame of a generation to be used as a starting point for the next clip (Storyboarding).
Combine Videos
Merges multiple generated clips into a single sequence.
Interpolation
Increases frame rate for smoother motion.
Upscale Video
Increases the spatial resolution of the final video output.
Description
Removed Image Filter Adjustments, Color Match, and some other nodes from the workflow.
Batch Color Corrector (beta) is now doing the heavy lifting in place of those nodes.
You can delete or bypass this node if it's not your cup of tea.
Added a Load VAE and Load CLIP node in case anyone is using a model that does not have these baked-in.
The noodles will have to be connected manually since using switches for these caused some issues that I spent the past 8 hours troubleshooting.
TL;DR regarding switches is that using switches for these can affect the output resulting in severely degraded results. YMMV.
FAQ
Comments (17)
New version of the 2.2 workflow uploaded.
For 5 second 480x640 videos: It takes about 44 seconds per KSampler for me on a 5060ti 16GB and 64GB RAM.
This is mostly due to being on pytorch 2.9.0 and cuda 13 and using the latest nvidia driver.
I came across this while checking on issues for comfyui on their project page.
I don't really recommend doing this unless you are comfortable doing that sort of thing and are willing to deal with the potential headaches of something breaking.
It took me a few tries to get it working. The best approach seems to be to fully uninstall torch (torch, torchvision, torchaudio) from your comfyui's python_embedded and then reinstall using pytorch 2.9.0 and cuda 13.
NGL I had to use Google's AI mode to walk me through it.
Thank you for mentioning me and my checkpoint!
Thanks for the great checkpoints!
I like the workflow. Could you add a feature to automatically calculate the resolution?
Thanks. That's probably beyond me at this time, but sounds useful.
I think I have this figured out. It's not pretty looking, but basically I use ComfyMath to convert the image resolution based on a target value entered.
Let's say we are going to convert an image from 1664x2432 to a 480x<calculated number>.
You would enter 480 into the Target Value and the math needed would be done automatically.
The important part here is people would need to understand that they need to toggle the boolean switch correctly depending on the image they loaded.
Setting it true will assume they are doing a vertical video with a width of 480, while setting it to false will assume a horizontal video with a height of 480.
I just tested this out as far as the numbers go, but putting into practice will be the next step.
Uploaded v3 to include this.
@Gladas Thanks for the update. I actually found a custom node called 'Resolution-Master', Can you check on that?
@g1263495582 I can check later. I'm hesitant to add more custom nodes usually because many end up being incompatible with one thing or another on my comfy install. Will be at least 6 hours before I can check.
Seems to work. ComfyUI-Manager didn't want to install it initially. I'll probably add it to the workflow later. Testing out some other stuff before I upload a new version.
@Gladas If a LoRA has a trigger word, do I need to change the CLIP connect pipe from Checkpoint Loader --> Prompt to LoRA Loader --> Prompt ?
@g1263495582 not sure tbh. I would say try it as-is and if doesn't seem to work, then adjust. I was able to use loras with how it is, but YMMV.
Hey! Worklow looks super nice, thanks for putting that together!
Although I can do only one generation, the second time my computer will freeze almost completely to the point i have to reboot
I suspect the RAM and VRAM, as they looked maxed out at the end of the first generation, and they dont seems to be cleaned out
How come you didnt need any clear cache nodes, am I missing something?
I tried those and they don't really seem to do anything (for me). You can try adding them and see if it helps you.
I know that there's an ongoing issue (last I checked) with memory and certain versions of pytorch. The comfy devs patched in a workaround from my understanding, but not sure if that's what is affecting you. This was one of the threads where they were talking about it: https://github.com/comfyanonymous/ComfyUI/issues/10355
@Gladas ah damn, i'll investigate. Thanks for your quick answer
Love your nickname btw
Tried to go for a build that'd match yours and now everything is completely broken T-T
Should have listened to your comment
"I don't really recommend doing this unless you are comfortable doing that sort of thing and are willing to deal with the potential headaches of something breaking."
@NerdyPixel Yeah, I was fighting with that for several hours. I get that sense of dread whenever I update comfy or any of the python dependencies. "is this gonna break my install??"
I hope you're able to get it resolved.