How do I turn one picture into a short AI video?

1. 01 / processPicture

   Authorised starting frame

2. 02 / processMotion

   One clear action in text

3. 03 / processGenerate

   Wan2.2 5B or hosted model

4. 04 / processExport

   Inspect the saved video

An image-to-video model uses your picture as the starting frame and generates the frames that follow. For a free local route, use the native Wan2.2 5B template in ComfyUI. For a hosted alternative, Runway Gen-4.5 accepts an image and a motion prompt. Neither route reproduces a particular dance from a reference video: that is character motion transfer.

Evidence: sources and the upstream template were inspected on 11 October 2026. We have not generated or compared videos with these routes. The diagram explains the process; it is not a result.

Choose the route before downloading

Local ComfyUI gives you the workflow and files without a per-generation provider charge. You supply the computer, electricity, storage and installation time. Start with your PC's memory and model requirements; a smaller parameter count does not make every video workflow lightweight.

Comfy documents native offloading for its 5B workflow and says it should fit on 8GB VRAM. The model publisher's separate command-line example specifies at least 24GB. These are different execution paths, not conflicting promises that either will work on your PC. We have not measured our own memory use or generation time. Comfy tutorial, publisher model card.

Runway avoids installing video models locally but uploads your picture to a provider and consumes credits. Its documented Gen-4.5 image-to-video route requires Standard or higher. Free Comfy instructions and Runway's own documentation already exist; this guide's contribution is connecting the right template, inputs, export and checks.

Prepare one simple input

Use a picture you own or have permission to animate. Save a PNG or JPEG without overlays obscuring the subject. For a first attempt, choose one clear subject with space around it and one modest action. A scene containing tiny faces, crowded hands and several moving objects makes failures harder to diagnose.

Decide on landscape or portrait before you start. Avoid an accidental crop that removes the very object you want to move. Keep the untouched original separately and give the working copy a simple name such as input.png.

Our proposed first experiment is a toy robot turning its head, with a stationary camera. That is a test brief, not an output we have made. Use your own matching picture rather than copying an unrelated example prompt literally.

Free local workflow: Wan2.2 5B

1. Install ComfyUI using its official installation instructions. Update it, then open the Template Library and search Wan2.2 5B. Select the native video generation template. Alternatively open the tutorial's Download JSON link. Use the actual upstream template, not an Animate replacement graph.
2. Obtain the exact model files linked in the official tutorial or template and put them in the following locations. Model weights are not included in our download. Restart or refresh ComfyUI so the loaders can see them.

ComfyUI/models/diffusion_models/wan2.2_ti2v_5B_fp16.safetensors
ComfyUI/models/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors
ComfyUI/models/vae/wan2.2_vae.safetensors

3. Match the three files in Load Diffusion Model, Load CLIP and Load VAE. Enable the bypassed Load Image node with Ctrl+B, upload input.png, and check its connection to Wan22ImageToVideoLatent. Without the active picture input, this hybrid template can run as text-to-video instead.
4. Keep the template's sampler defaults for the first attempt. The inspected graph records 20 steps, CFG 5, uni_pc, simple, denoise 1 and model shift 8. Its node metadata records core version 0.3.45; that identifies the saved graph and does not guarantee compatibility with every present installation. Record the ComfyUI version you actually run.
5. Check size and length in Wan22ImageToVideoLatent. The inspected template has 1280 × 704, 121 frames and batch size 1. Create Video uses 24fps: 121 ÷ 24 is approximately 5.04 seconds of encoded frames. These are upstream defaults, not our recommended settings for an unmeasured 8GB computer. If memory is tight, use a lower compatible size and shorter frame count accepted by the node, and record what you changed.
6. Replace the positive prompt with a short description matching your subject and desired movement. Retain the provided negative prompt for the first comparison. Press Run and wait for completion. Check the error panel if no output appears.
7. Inspect Create Video → Save Video. The graph's filename prefix is video/ComfyUI, with format and codec set to auto; save/download the result from its output preview or find the saved video in ComfyUI's configured output directory. If you need an MP4 for a platform, select MP4 where offered and reopen that exported file before continuing. Do not assume auto has produced a particular container. Upstream graph, native video saving implementation.

A motion prompt to start with

The small toy robot slowly turns its head to the left and pauses.
Its body stays in place. The camera remains still.

This is our untested starter prompt. Change “toy robot” to the subject actually present. Requesting one action gives you something observable to assess. Keep the original input, prompt and seed; change one thing at a time when comparing attempts. A fixed seed helps record an experiment, but is not a guarantee of the same character or result across changed versions.

Paid hosted workflow: Runway Gen-4.5

1. In Runway, open Apps View, search Gen-4.5, or choose it in Tool Mode with Video selected. Check your plan and the generation charge before proceeding.
2. Drag the authorised picture into the prompt window. Describe the motion, then confirm the aspect ratio and crop. Choose a short duration; the documented range is 2–10 seconds. Select 24 or 25fps in Advanced settings.
3. Generate, review the result, and use the download control below it. Reopen the downloaded file and check the beginning, middle and end. Record unsuccessful attempts too.

Runway lists 12 credits per second for Gen-4.5: a four-second generation is 48 credits before optional extra processing. Higher-tier ProRes/PNG output adds 5 credits per second; ordinary video download is sufficient for this first exercise. Subscription price, tax and your account's purchased-credit cost were not verified here. Gen-4 Turbo is a separate older option with lower documented credit use; it is not the same model or a tested quality comparison. Gen-4.5 instructions, Gen-4/Turbo details.

Check the result and fix the right failure

If the clip ignores the picture, check that Load Image is enabled and connected. If a model dropdown is empty, compare the filename and folder, not merely the download's completion indicator. If nodes are missing, follow the missing-node diagnosis; do not install an unrelated pack to replace a native node.

For distorted movement, simplify the action and compare the same input. For an out-of-memory error, record the failing dimensions, frame count and memory before reducing them. Slowing the saved video changes playback speed; it does not add better generated motion.

Inspect subject identity, unwanted camera movement, flicker and cropped limbs. This graph has no audio input connected, so our local route does not generate speech or a soundtrack. Add authorised audio in a separate editor if wanted; use the lip-sync guide when the mouth must follow spoken words.

Download and reproduction record

The starter pack contains our prompt, checklist and reproduction log. It contains no weights or executable graph. Get the current official graph from its linked source, retain its licence with any redistribution, and note any changes. The Wan model card is labelled Apache-2.0; your input and added audio still need their own permissions.

Record input hash, model filenames, software version, dimensions, frames, fps, prompt, seed, time, failed attempts and real exported filename. Generation: UNTESTED. Until an actual run is retained, output quality, speed, hardware compatibility and cost per usable clip remain unknown.

The small toy robot slowly turns its head to the left and pauses. Its body stays in place. The camera remains still.