What kind of image works best as input?

Sharp, well-lit, single-subject images. The model animates what it sees, so a cluttered photo produces cluttered motion. Aim for:

• At least 1024px on the long edge.
• One clear subject, ideally centered.
• Consistent lighting without harsh mixed color.
• No heavy compression artifacts or watermarks.

Portraits and product shots animate well. Wide landscapes with fine texture (forests, crowds) tend to shimmer.

How do I set the motion in Wan 2.2?

Your prompt describes movement, not appearance. Do not re-describe the subject — the image already carries that. Write things like “slow push toward the subject, gentle head turn, hair drifting” or “camera pans right, clouds move slowly.” Keep it to one dominant motion. Two competing moves produce a muddled result.

There is usually a motion-strength or motion-amount control. Start low. A subtle 0.5-style setting reads as natural; cranking it up makes the whole frame warp.

Why does my output look like it is warping?

Three causes, in order of frequency. First, motion strength set too high — the model distorts geometry to invent movement. Second, a source image with busy texture, which the model tries to animate everywhere at once. Third, frame count too high for the motion you described, so the model fills the extra seconds by drifting. Fix them by lowering strength, simplifying the image, and cutting the clip shorter.

How long should the clip be?

Four to five seconds covers most needs and keeps VRAM in check. Longer clips on image-to-video often develop a slow zoom-out or identity drift, because the model has to keep inventing motion that was not implied by the still. If you need a longer shot, render two shorter ones from slightly different crops and cut between them.

Can I chain it with text-to-video?

Yes, and it is a smart workflow. Generate a keyframe in text-to-video, export that frame as a still, then run it through image-to-video for the animated portion. That gets you precise framing from the first pass and stable motion from the second. Compare that control with a hosted tool such as Pika 3, which trades some of it for a faster, more stylized path.

Quick Summary

• Wan 2.2 image-to-video animates a still and preserves its composition.
• Use sharp, single-subject images at 1024px+; avoid busy textures.
• Describe motion in the prompt; do not re-describe the subject.
• Keep motion strength low and clips to 4-5 seconds to avoid warping.
• Chain text-to-video keyframes into image-to-video for fine control.

What is Wan 2.2?

It is a text-to-video and image-to-video model released as open weights, trained to handle realistic motion and camera movement. The smaller 5B variant targets consumer cards; the 14B variant wants more VRAM but produces cleaner detail. Because it is open, you can also fine-tune it on your own footage, which closed services do not allow.

What hardware do I actually need?

You have options depending on VRAM:

• 8GB (e.g. RTX 3060 Ti): the 5B model at 480p, with offloading enabled.
• 12GB (RTX 3060 12GB): 5B at 720p is workable.
• 24GB (RTX 4090): the 14B model at 720p runs comfortably.

System RAM matters too — 32GB is a sane minimum, because weights spill over when VRAM runs out.

How do I install and run it?

1. Install Python 3.10+ and CUDA drivers for your card.
2. Set up a venv and install PyTorch for your CUDA version.
3. Grab the Wan 2.2 weights from the official model repo.
4. Install ComfyUI — it has first-class Wan support through official workflow templates.
5. Load the provided Wan text-to-video template and swap in the checkpoint.
6. Set resolution, frame count, and seed, then queue.

Wan ships reference workflows, so you do not build the graph from scratch. Start with the template and adjust one setting at a time.

Is Wan 2.2 really free?

The weights are free to download and run, and there is no metered fee. Your real cost is the GPU and electricity. That is a genuine advantage over hosted tools that charge per second of video — if you render often, the hardware pays for itself. The trade-off is time: local renders are slower than a cloud service on equivalent quality, and you handle all the troubleshooting.

Wan 2.2 vs hosted generators

Against Seedance 1.5, Wan 2.2 loses on native audio and out-of-box polish but wins on cost-per-clip and fine-tuning freedom. Against Hunyuan Video 2, the two are close; Wan’s smaller 5B model runs on weaker GPUs, which makes it the friendlier entry point for a mid-range card.

Quick Summary

• Wan 2.2 is Alibaba’s open-weight video model with 5B and 14B variants.
• An 8GB card runs the 5B model at 480p; 24GB runs 14B at 720p.
• Install via ComfyUI with the official Wan workflow templates.
• No per-clip fee — you pay in hardware and render time instead.
• Beats hosted tools on cost and customization, loses on audio and convenience.