OmniFit logo OmniFit Blog home
ComfyUI workflow • Wan 2.1 example

How to make an AI video in ComfyUI with Wan 2.1

If you want one ComfyUI video workflow that feels practical instead of overwhelming, start with Wan 2.1 image-to-video. The official ComfyUI tutorial already breaks it into a small native graph with a start image, a prompt box, a few model loaders, and one video node, which makes it a strong first example for creators who want a repeatable setup.[2]

đź“… Published: October 23, 2026
Beginner-friendly example
Image-to-video
Official ComfyUI workflow
Official Wan 2.1 image-to-video workflow preview from the ComfyUI documentation
Best for

Use this when you already have one strong still frame and want to turn it into a short AI video clip without building a giant graph first.

Why this example

The official tutorial exposes the core pieces clearly: diffusion model, CLIP, VAE, CLIP Vision, input image, prompt, and one Wan image-to-video node.[2]

What makes it good

It is small enough to understand in one sitting, but real enough to teach how ComfyUI video graphs are wired.

What the docs confirm

ComfyUI ships both 480p and 720p Wan 2.1 image-to-video variants, and the official steps are nearly identical across both versions.[2]

What to watch for

Your result quality depends heavily on the starting image. ComfyUI can animate a weak frame, but it cannot make a weak composition feel intentional.

Companion watch

One YouTube walkthrough to keep open while you build

Arcane Ai Alchemy’s tutorial is a strong companion because it focuses on the native Wan 2.1 ComfyUI image-to-video workflow, walks through the nodes, and calls out practical controls like prompting, camera motion, troubleshooting, and video length. The video description also lists timestamps for models, nodes, prompting, and camera movement, so it is easy to jump to the part you need.[5]

Arcane Ai Alchemy 111K+ views 24:35
Step-by-step tutorial

A simple ComfyUI AI video workflow to copy first

A lot of ComfyUI video posts lose beginners because they start with giant graphs, many custom nodes, or too many quality hacks at once. Wan 2.1 image-to-video is a better first example because the official ComfyUI docs already present it as a native workflow with a clear input image, a prompt, a small model stack, and a direct run loop.[2] That makes it easier to understand what each part does before you add more control.

The core idea

Do not ask ComfyUI to invent the whole film from nothing on day one. Start from one still frame you already like, then use image-to-video to add motion. That is the easiest way to learn how video graphs behave without also fighting scene composition at the same time.

The 6-step workflow

01

Pick one strong start image first

The official workflow begins with a LoadImage node, which tells you something important: this example assumes the shot already exists before motion starts.[2] That is why this is such a good beginner example. You can focus on animation, not on figuring out both composition and motion at the same time.

  • Choose a clean still with one clear subject.
  • Avoid crowded frames with many tiny moving parts.
  • Start with a shot that would already work as a poster frame.
Default Wan 2.1 start image from the ComfyUI workflow template
The official template ships with a default start image for the LoadImage node, which is a good reminder that this workflow is image-led first, motion second.[2]
02

Load the four required model pieces

The ComfyUI docs list four required loaders for the Wan 2.1 image-to-video graph: diffusion model, CLIP, VAE, and CLIP Vision.[2] In plain English, that means the workflow needs one model for the generation itself, one text encoder for your prompt, one VAE for image/video latent conversion, and one vision encoder so the graph can understand your input frame.

  • Load Diffusion Model: Wan 2.1 image-to-video checkpoint.
  • Load CLIP: text understanding for the prompt.
  • Load VAE: latent/image conversion.
  • Load CLIP Vision: visual understanding of the start image.

For the official 480p example, the docs name wan2.1_i2v_480p_14B_fp16.safetensors, umt5_xxl_fp8_e4m3fn_scaled.safetensors, wan_2.1_vae.safetensors, and clip_vision_h.safetensors explicitly.[2]

03

Start with the 480p workflow before jumping to 720p

The official docs provide both 480p and 720p versions of the same Wan 2.1 image-to-video workflow and note that Wan separates these models by resolution.[2] For a first run, 480p is the better learning path because it lets you validate the graph and prompt logic before you spend more time and memory on higher resolution.

  • Run 480p first to confirm the graph works end to end.
  • Only switch to 720p after you like the motion behavior.
  • Treat 720p as the “cleaner second pass,” not the first experiment.
Official ComfyUI Wan 2.1 image-to-video step diagram
The official flow diagram is useful because it shows the workflow as a finite setup process rather than a mystery graph.[2]
04

Write a prompt that describes motion, not just style

The docs treat the prompt inside the CLIP Text Encoder as optional, but in practice it is where you tell the graph what kind of movement should happen.[2] A weak prompt only repeats the image. A better prompt explains the motion: slow push-in, wind through hair, subtle head turn, cloth movement, camera pan, or environmental movement.

  • Describe motion verbs, not only the scene look.
  • Prefer one clear movement direction over five competing actions.
  • Keep the motion matched to the subject and framing you already chose.

The companion YouTube tutorial spends dedicated sections on prompting, camera control, subject focus, prompt troubleshooting, and video length, which is exactly why it pairs well with the official docs.[5]

05

Adjust dimensions only after the first clean run

The official WanImageToVideo node allows dimension changes if needed.[2] But for a first test, keep the template close to default. New ComfyUI users often break working graphs by changing too many variables at once. First prove the graph runs. Then tune dimensions, framing, or prompt strength one change at a time.

  • Do one clean default run first.
  • Then change only one thing: prompt, dimensions, or seed behavior.
  • Keep a note of what changed so you can repeat good results.
06

Queue the workflow, review the clip, then iterate upward

The docs end with the most basic instruction—click Run or use Ctrl(cmd)+Enter to execute the generation.[2] That sounds simple, but the real creator lesson is what comes next: review the short clip, decide whether the motion feels believable, then iterate from there. Do not make the graph bigger until the basic motion already works.

  • Check whether the motion is believable, not just “animated.”
  • If the shot feels unstable, simplify the prompt.
  • If the motion is good, then try 720p, first-last-frame, or consistency workflows next.[3][4]
The easiest way to learn ComfyUI video is not to chase the biggest graph. It is to make one still image move in a believable way, then build from there.

Why this example works so well for beginners

ComfyUI can look intimidating because the interface makes every dependency visible. But that is also the advantage. In this Wan 2.1 example, you can literally see the job split: one image comes in, model loaders prepare the graph, one text encoder adds the motion intent, and the Wan video node turns that into a clip.[2]

That is easier to reason about than a tool that hides everything behind one button. Once you understand this graph, the next official directions—text-to-video, first-last-frame, and character-consistency workflows—start feeling like extensions instead of totally different worlds.[1][3][4]

Tool stack, in plain English

Input image

Your visual anchor. This is the shot the whole clip grows from.

Wan 2.1 model stack

Diffusion model, CLIP, VAE, and CLIP Vision together form the minimum native stack for this workflow.[2]

Prompt box

Use it to tell the graph how the shot should move, not just what the shot looks like.

Optional next step

After this workflow clicks, move to text-to-video, first-last-frame, or character-consistency variants for more complex shots.[1][3][4]

Easy mistake to avoid

Do not treat a higher resolution as a better starting point. A clean 480p test teaches you more than a broken 720p run. Get the motion logic right first, then spend more compute on the nicer version.[2]

Sources

[1] ComfyUI docs — Wan 2.1 text-to-video

Used for the “what to try next” context: ComfyUI has a neighboring official Wan text-to-video path, which helps frame image-to-video as one entry point rather than the whole video story.

Read source
[2] ComfyUI docs — Wan 2.1 image-to-video

Main source for the workflow choice, required models, 480p/720p split, input image, optional prompt field, optional dimension edits, and the official run steps.

Read source
[3] ComfyUI docs — Wan first-last-frame workflow

Used for the “grow from here” recommendation. Once the basic image-to-video graph makes sense, first-last-frame is a natural next upgrade for more controlled shot endpoints.

Read source
[4] ComfyUI docs — Wan character consistency

Used for the advanced-next-step note. It shows that ComfyUI’s official Wan docs also branch into consistency workflows after the simpler entry workflows.

Read source
[5] YouTube — Arcane Ai Alchemy

Supporting tutorial video: ComfyUI - Wan 2.1 Image to Video, Made Simple. Used for the embedded walkthrough and for the note that the video covers nodes, prompting, camera control, troubleshooting, and video length.

Watch on YouTube