Several months ago, I shared a Wan-based video generation workflow that pulled together multiple control mechanisms for the model into a common setup, and it received a very enthusiastic response from folks who follow my work.
This could be considered a spiritual successor to that post, although the video generation landscape has changed so much that it’s also a very different beast.
This time it brings together 4 different ways the newer, remarkably capable Wan2.2 open-source models can be used into a single modular ComfyUI workflow:
- Image Generation
- Text-to-Video
- Image-to-Video
- First-Frame / Last-Frame
As it’s often the case, the workflow emerged from dissatisfaction with existing solutions.
The result looks deceptively compact – in part thanks to SubGraphs, quite possibly the best ComfyUI feature ever released – but it’s quite versatile and packs a serious punch.
Some technical details (the workflow contains a more detailed documentation node):
- For best video results use 16:9, 9:16 or 1:1 aspect ratios with resolutions up to 720p (1280×720, 720×1280, 640×360, 768×768, 512×512, etc.)
- Still image generation works fine at higher resolutions (e.g. 1920×1080).
- In addition to a video, it also outputs an image batch of individual frames, making it possible to upscale the frames and combine them into truly high-resolution video.
- 16 and 24 fps are optimal framerates, but I have had excellent 30 and even 60 fps results, although things get VERY resource-intensive at those settings.
- It uses ComfyUI’s repackaged fp8_scaled model variants, making it reasonably usable on lower VRAM systems (although a minimum of 24GB VRAM is still highly recommended).
- The FrameSnap nodes in the workflow ensure that the number of frames set are adjusted to comply with Wan’s “4n+1” temporal compression and frame architecture. This eliminates issues like truncated clips or weird visual glitches.
The best part? It’s absolutely FREE, no string attached.
Just grab it and send some good vibes my way: 4-In-Wan Workflow.zip
I realize there are additional members of the Wan model family – like Audio-To-Video and Character Animate / Replacement – but those workflows are different enough that attempting to incorporate them here would have made things considerably more convoluted as well as resource-intensive.