Animating Stills in Practice: Wan 2.2 Image-to-Video

Updated September 12, 2026

Image-to-video is the step that turns a gallery into a film. You already have a still you like — a trained character, the right pose, the right light — and you want it to move without losing the face, the outfit, or the scene. Wan 2.2 I2V (image-to-video) is the model we reach for, and this is what we’ve learned running it in FlixML.

What I2V does

You give it one image and a motion prompt. It generates a short clip — a handful of seconds — where the first frame is anchored to your still and the rest is inferred motion. It is not a filter over your image and it is not interpolation between two frames. It’s a video model conditioned on your still, which is why it can add camera movement, hair and cloth motion, and subtle expression change while keeping the subject recognizable.

The mental model that keeps you out of trouble: the still sets the identity, the prompt sets the motion. Ask for what should move, not what should appear — the appearance is already decided by the image you fed it.

The knobs that matter

• Steps. More sampling steps buy coherence up to a point, then just cost time. Around 20 is a sane working point for I2V; push higher only when you see temporal jitter you can’t prompt away.
• CFG (guidance). Controls how hard the model follows your motion prompt. Low CFG (~3–4) tends to give natural, believable motion; crank it and motion gets literal and stiff, and artifacts creep in. When a clip looks “possessed,” drop CFG before you touch anything else.
• Clip length / frame count. Longer clips drift — the subject slowly stops looking like itself. Generate shorter and string clips together into a scene rather than asking for one long take. This is also how you keep each shot cheap to re-roll.
• Resolution. I2V is VRAM-hungry. Match the resolution to what your node can hold; a clip that OOMs at 1080p renders fine downscaled, and you lose less than you’d think once it’s in motion.

Where it breaks (and the fix)

• Face drift on long clips. The subject melts the further from frame one you get. Fix: shorter clips, then concatenate. Don’t fight physics with steps.
• “Too much” motion. A busy motion prompt plus high CFG produces flailing. Fix: describe one clear action, lower CFG.
• Static output. If nothing moves, your prompt is describing the image, not the motion. Fix: prompt verbs — “slowly turns her head,” “wind moves the coat,” “camera pushes in.”
• OOM on the node. Video needs more headroom than images. Fix: put video on your highest-VRAM node and drop resolution before you drop quality settings.

The workflow that makes it repeatable

The reason to do this inside FlixML rather than by hand: I2V is one API call downstream of the image you already generated. Same character, same pipeline — generate the still, then animate it, then chain the clips into a scene, all as ordinary async jobs your agent submits and reconciles. The individual clip is where the craft lives; the assembly is where the automation pays off.

Wan 2.2 I2V docs → · Workflow catalog →