Image-to-Video: Animating a Still Image

From Still to Moving

Image-to-video (I2V) takes a static image — a photo, an illustration, an AI-generated image — and generates a short video that starts from that exact frame and animates it forward in time. The model synthesizes plausible motion: a portrait might blink and turn its head, a landscape might show wind moving through trees, a product shot might slowly rotate.

This is distinct from text-to-video because the first frame is anchored to your input image. The model is not free to invent the scene — it must start exactly as your image appears and transition from there. This makes I2V ideal for animating specific AI images you have already generated, or for bringing reference photos to life.

Wan 2.2 I2V Workflow

The Wan 2.2 I2V workflow is built into ComfyUI. Load it from the Workflow Templates panel (select "Wan 2.2 Image-to-Video") or download the workflow JSON from docs.comfy.org. The key difference from the T2V workflow is one added node:

• Load Image: loads your starting frame. This image becomes the first frame of the generated video.
• WanVideo Loader: now select the I2V model variant (wan2.2_i2v_480p_14B.safetensors or similar). The I2V model is a different model file from the T2V model — both need to be downloaded separately.
• CLIPVisionEncode: encodes the reference image for visual conditioning alongside the text prompt. The image encoding and text encoding both feed into the WanVideo Sampler.

The CLIP Text Encode (Wan) node in I2V mode acts as a motion prompt — describe what you want to happen in the animation, not what the scene looks like (the model already knows from the image). "The woman slowly turns to look at the camera," "the water begins to ripple," "a gentle breeze moves through the leaves." Keep the motion prompt focused on action rather than appearance.

Image Preparation for I2V

The quality of your starting image significantly affects I2V output:

• Use images at the resolution the model was designed for — Wan 2.2 I2V produces 480p (832×480 or 480×832 portrait) by default. Crop or resize your image to match before loading.
• Clean, well-composed images with clear subject-background separation animate better than busy, complex scenes
• Images generated by FLUX or SDXL work well as I2V inputs — they have the resolution and detail quality that makes animation coherent
• Real photographs can also be used, but may require cropping to the model's expected aspect ratio

Use a Resize node or the built-in image resize in Load Image to match the target resolution before passing to CLIPVisionEncode and the WanVideo Sampler.

HunyuanVideo I2V Alternative

HunyuanVideo also supports image-to-video and generally produces smoother motion than Wan 2.2 I2V for portrait and character animation. If you have enough VRAM for HunyuanVideo (discussed in Lesson 7), it is worth comparing results for character-focused I2V work. The workflow structure is similar — Load Image feeds into a visual conditioning node before the sampler.

First-Last-Frame Interpolation (FLF2V)

Wan 2.2 includes a third I2V variant: FLF2V (First-Last-Frame to Video). You supply both a starting image and an ending image, and the model generates the transition between them. This is powerful for controlled motion:

• Start frame: person standing still
• End frame: person mid-stride
• Output: smooth walking motion between the two states

FLF2V requires two Load Image nodes (one for start, one for end), both feeding into the conditioning. The motion between frames is synthesized entirely by the model — you cannot control the intermediate path, only the endpoints. Use this for transitions where you know exactly where you want to start and end but want the model to handle the in-between motion naturally.