Wan 2.2 Image-to-Video Spicy API

Wan 2.2 Image-to-Video (Tongyi Wanxiang Wan 2.2 Image-to-Video) is Alibaba Cloud's high-definition video generation model that brings static images to life as fixed 5-second videos based on a reference start frame and an optional text prompt. The Wan 2.2 series was open-sourced on July 28, 2025 as the first video diffusion model family to introduce a Mixture-of-Experts (MoE) architecture: a high-noise expert handles overall layout in the early denoising stages while a low-noise expert refines details later — roughly 27B total parameters with only 14B active per step, significantly improving quality at nearly unchanged inference cost. Training data grew 65.6% more images and 83.2% more videos over Wan 2.1, and on the official Wan-Bench 2.0 the models surpass leading commercial models across most key dimensions. The I2V A14B model is purpose-built for image-to-video, with reduced unreasonable camera movement and better support for stylized scenes.

The model is offered through the iCreat platform as an API using a two-step asynchronous task flow: submit a task to obtain a task_id, then poll the result endpoint until a terminal status. Output duration is fixed at 5 seconds with 480P and 1080P resolution tiers; the aspect ratio automatically follows the input image. Billed by output duration: $0.02 per second at 480P and $0.10 at 1080P.

Model Positioning

Wan 2.2 Image-to-Video targets "bringing static assets to life in HD." It is a dedicated tool for turning an existing image into a 5-second motion video: e-commerce product shots into showcase clips, portraits into subtle motion pieces, illustrations and design drafts into dynamic demos — all from one start-frame image plus an optional English prompt. Within the Tongyi Wanxiang family it is the 2.2-generation I2V tier: sharing the MoE architecture and aesthetic training with the text-to-video A14B, purpose-optimized for image-to-video with steadier camera work and friendlier stylized-scene support. The fixed 5-second duration and two-tier resolution make it a cost-controlled, consistency-friendly choice for batch image-to-video pipelines.

Core Capabilities

Stable Animation Driven by the Start Frame

Generates a 5-second video anchored to the reference start frame: subject appearance and composition are locked by the image, and the prompt only needs to describe motion, camera, and dynamics — output stays highly faithful to the original.

MoE Two-Expert Architecture

A high-noise expert handles early-stage overall layout while a low-noise expert refines late-stage detail — about 27B total parameters with only 14B active per step, improving quality at nearly unchanged inference cost.

Cinematic Aesthetics

Trained with meticulously curated aesthetic data carrying fine-grained labels for lighting, composition, and color, enabling more precise and controllable cinematic-style generation.

Complex Motion and Generalization

Training data up 65.6% more images and 83.2% more videos over Wan 2.1, markedly improving generalization across motion, semantics, and aesthetics.

Negative Prompts

negative_prompt excludes unwanted elements (low quality, blur, facial distortion, extra limbs, jittering, artifacts — up to 500 characters), keeping output quality predictable.

Pricing

Total Cost = Unit Price × Output Video Duration (fixed at 5 seconds). Only the output duration is billed; the costUSD field in the response returns the actual cost once the task succeeds.

Application Scenarios

• E-commerce product clips: turn static product shots into 5-second showcase videos in one step
• Portrait and photo motion: natural head turns, smiles, and hair movement from still photos
• Illustration and design demos: dynamic presentation of static visual drafts
• Brand asset activation: posters and key visuals converted to short motion placements
• Storyboard previews: camera-movement previews from concept frames

Model Comparison

Same-Series Comparison

Cross-Model Comparison

Same-series data comes from Tongyi Wanxiang's official releases; this model's specifications follow the iCreat channel documentation.

Why Choose Wan 2.2 Image-to-Video?

• One start-frame image plus one prompt is all it takes — dramatically higher reuse of static assets
• $0.02 per second at 480P and just $0.10 per 5-second video — extremely low cost for batch animation
• MoE two-expert architecture plus cinematic aesthetic data delivers outstanding quality and stability at this price
• Aspect ratio automatically follows the input image — no extra configuration
• Negative prompts filter common flaws, keeping output quality predictable

Specifications

Architecture

Wan 2.2 is the Tongyi Wanxiang video generation series open-sourced on July 28, 2025, building on Wan 2.1 with three key upgrades: Mixture-of-Experts — the first MoE introduced into video diffusion models, with the A14B series using a high-noise/low-noise two-expert design switched by signal-to-noise ratio, totaling about 27B parameters with 14B active per step; upgraded training data — 65.6% more images and 83.2% more videos, markedly strengthening motion, semantic, and aesthetic generalization; and the high-compression Wan 2.2-VAE — enabling the 5B model to generate 720P video at 24fps on consumer-grade GPUs. The I2V A14B is the series' MoE model purpose-optimized for image-to-video, supporting 480P and 720P (the channel additionally offers a 1080P tier), with reduced unreasonable camera movement. The series is open-sourced under Apache 2.0.

Notes

• Submit and query must be chained with the same task_id
• While processing, result is []; keep polling, with a suggested interval of 2–5 seconds
• FAILED is terminal; check request parameters and the start frame before resubmitting
• Video duration is fixed at 5 seconds: duration, prompt_extend, and seed parameters are not supported
• img_url accepts a public URL or Base64; mind the 20MB file-size cap
• The start frame must be 240–8000 px per side with an aspect ratio between 1:8 and 8:1
• prompt must be in English, recommended within 800 characters
• The aspect ratio automatically follows the input image; resolution only selects the clarity tier
• input and parameters are sibling top-level fields; do not nest them
• costUSD is returned only when the task succeeds; failed tasks do not produce a cost field

FAQ

How do I get started with the Wan 2.2 Image-to-Video API?​

Register on the iCreat platform and obtain an API Key from the console, send a generation request to the submit endpoint, then poll the result endpoint with the returned task_id. All requests are authenticated with the Authorization: Bearer header. Keep your API Key safe and never expose it in client code or public repositories.

How is the cost calculated?​

Total Cost = Unit Price × 5 seconds (fixed duration). A 480P video costs 0.02 × 5 = $0.10; a 1080P video costs 0.1 × 5 = $0.50. For example, batch-generating 100 480P product clips costs $10. The costUSD field in the response gives the actual cost once the task succeeds.

What are the requirements for the start-frame image?​

JPG, JPEG, PNG, BMP, or WEBP via a public URL or Base64; each side must be 240–8000 px, the aspect ratio between 1:8 and 8:1, and the file no larger than 20MB. The output video's aspect ratio automatically follows the start frame — no need to specify it separately.

How should I write the prompt?​

Use prompt in English to describe motion, camera work, and dynamics only (the image already locks subject and composition), e.g. "Gentle breeze rustles the hair, the character turns toward the camera with a subtle smile" — recommended within 800 characters. Pair it with negative_prompt to exclude low quality, blur, distortion, and jittering for steadier output.

Can I change the video duration?​

No. This model outputs a fixed 5 seconds and does not support the duration parameter (prompt_extend and seed are also unsupported). For variable-duration image-to-video, use Wan 2.6 Reference-to-Video (4–15 seconds) or MiniMax H3 Max Image-to-Video (4–15 seconds).

What if the task fails?​

FAILED is a terminal status. Check in order: whether img_url is a valid public URL or Base64, whether the image format and dimensions are within range, whether resolution is 480P or 1080P, and whether unsupported parameters like duration were passed by mistake. Fix any issue and resubmit. The cost field is returned only when the task succeeds.