Wan 2.2 Fast API

Wan 2.2 Fast takes a still image and a sentence about what should happen, and returns a short video clip in which the picture moves: a fox lifts its head, a camera drifts around a product, wind catches a coat. It is built on Wan2.2-I2V-A14B, the image-to-video model Alibaba released with open weights on 28 July 2025, and it has been compressed by the optimisation company Pruna AI so that a clip comes back in under half a minute at 480p.

That combination is what makes it interesting this week. Wan 2.2 arrives as one of the strongest open video models anyone can download, with a new mixture-of-experts design and training data aimed squarely at cinematic looks. The full model, though, wants a data-centre GPU and patience. The fast version, launched on Replicate within two days of the open release, turns it into something you can call from an API and run hundreds of times in an afternoon.

On Unicorn API it is a single model, wan-video/wan-2.2-i2v-fast. It returns an 81-frame clip, about five seconds, at 480p or 720p, which makes it one of the quickest ways to animate an image anywhere in the catalog.

Alibaba's Wan team and the open Wan 2.2 release

Wan, or Tongyi Wanxiang in Chinese, is the visual generation line of Alibaba's Tongyi Lab, the research group that also builds the Qwen language models. Its previous release, Wan 2.1, was open-sourced in February 2025, and by the time Wan 2.2 shipped Alibaba counted more than 5.4 million downloads of its Wan models on Hugging Face and ModelScope. Wan has become a staple for people who build their own video pipelines.

Wan 2.2 keeps that open approach. Alibaba published three models: Wan2.2-T2V-A14B for text to video, Wan2.2-I2V-A14B for image to video, and Wan2.2-TI2V-5B, a smaller model that handles both and fits on a consumer graphics card. All three carry the Apache 2.0 licence, which allows commercial use, and the weights sit on Hugging Face, GitHub and ModelScope. Alibaba presented the release as the world's first open-source video generation model with a mixture-of-experts architecture.

Wan 2.2 Fast is the image-to-video A14B model, not the 5B one. Replicate describes it as a very fast, Pruna AI-optimised version of Wan 2.2 A14B image-to-video. Pruna AI, a company that sells model compression, says it optimised Wan 2.2 the day after release and deployed it on Replicate, and announced the image-to-video endpoint on 30 July. Replicate greeted the arrival of Wan 2.2 with a blog post titled "Open source video is back".

What changed from Wan 2.1

Wan 2.2 is a bigger step than the version number suggests. Alibaba lists three main changes.

A mixture-of-experts design. Wan 2.1's 14B model was a single network. Wan 2.2's A14B models split the work between two experts of roughly 14 billion parameters each, 27 billion in total, with only one active at any denoising step. Alibaba says activating only one expert per step reduces computational consumption by up to 50 percent while adding capacity.

Cinematic training. The team curated aesthetic data with detailed labels for lighting, composition, contrast and colour tone. The practical result, according to Alibaba, is finer control over lighting, time of day, colour tone, camera angle, frame size, composition and focal length. Prompts that use film vocabulary, such as "low-angle shot", "golden hour" or "shallow depth of field", are meant to land more reliably.

More data, more motion. Wan 2.2 was trained on 65.6 percent more images and 83.2 percent more videos than Wan 2.1. Alibaba says the larger set brings better generalisation across motion, meaning and aesthetics, and better handling of large, complex movement.

For the image-to-video model specifically, Alibaba promises "more stable video synthesis with reduced unrealistic camera movements and enhanced support for diverse stylized scenes". In other words, less of the random camera drift that plagued earlier image animators, and better results when the input is an illustration or a painting rather than a photo.

On Wan-Bench 2.0, a benchmark Alibaba built itself, the company reports that Wan 2.2 beats leading closed commercial video models on most of the dimensions it measures; Pruna's launch post names Seedance, Hailuo, Kling and Sora among them. These are the maker's own tests, not an independent leaderboard. Artificial Analysis added Wan 2.2 A14B to its public Video Arena on release day, where user votes will decide how it really stacks up.

What "Fast" means

The full Wan2.2-I2V-A14B model needs at least 80 GB of graphics memory to run on a single GPU, and a standard run takes many denoising steps. Pruna applied compression from its own toolkit to cut that down. It has described its recipe for the Wan 2.2 endpoints as guidance and step distillation, the technique Black Forest Labs used to make FLUX Schnell, which lets the model produce a clip in far fewer steps and on a single GPU rather than several.

The numbers Pruna published at launch: a 480p clip from one image in 28 seconds and a 720p clip in 80 seconds on a single H100, which it puts at 1.5 to 2 times faster than the base endpoints. Pruna says the compressed model reaches "state-of-the-art quality comparable with Wan 2.2". It has not published side-by-side measurements, so treat that as a claim to test on your own images. Pruna's own disclaimer asks users to check results for their use case.

How the two experts share the work

A diffusion video model starts from noise and removes it step by step. Wan 2.2 hands the early, noisy steps to a high-noise expert that decides the overall layout: where things are and how they move. Once the signal-to-noise ratio crosses a set threshold, a low-noise expert takes over and refines detail such as texture, edges and faces. Each expert specialises in the stage it is good at, and because only one runs at a time the compute per step stays close to that of a 14B model. Alibaba reports that this two-expert split reaches a lower validation loss than the non-MoE baselines it tested.

Wan 2.2 Fast specs on Unicorn API

Length is set in frames rather than seconds. At the default 16 frames per second, 81 frames play for about 5 seconds and 121 frames for about 7.6. The upstream documentation notes that 81 frames give the best results. Raising frames_per_second without interpolation does not add frames; it plays the same frames faster, so 81 frames at 24 fps give a punchier clip of about 3.4 seconds. interpolate_output instead generates in-between frames after the fact, so the clip keeps its length but plays at a smoother 30 fps.

end_image is the most useful extra. Give the model the frame you want the clip to finish on, and it plans the motion between the two pictures, which makes transitions, reveals and before-and-after shots possible.

What Wan 2.2 Fast does well

Volume. With a 480p clip back in about half a minute, you can animate a whole product catalogue or try twenty versions of one idea in a single sitting.

Respecting the start frame. The image sets the look, and Wan 2.2 keeps it: colours, composition and character design stay recognisable as the scene moves, with less unprompted camera wander than earlier image animators.

Camera language. The cinematic training shows in how it responds to directions like push-in, orbit, crane up or handheld. A simple product still can become a slow reveal with a single sentence.

Natural, mid-sized motion. Hair and fabric in wind, flowing water and simple gestures come out fluid at five seconds.

Stylised inputs. Better support for stylised scenes is one of the specific improvements Alibaba claims for the image-to-video model, which makes illustrations, paintings and 3D renders a good fit.

Speed for iteration. Pruna measured about half a minute per 480p clip, which should make it practical to iterate on a prompt the way you would with an image model.

Where it falls short

Image to video only, and silent. There is no text-to-video mode on this route and no audio. Music, voice and effects have to be added afterwards.

Short clips. 121 frames is the ceiling, about 7.6 seconds at 16 fps, and the best results come at 81 frames. Five seconds is room for one camera move and one action; prompts that ask for three events tend to come out rushed and smeared.

16 fps by default. Wan 2.2's A14B models generate at 16 frames per second, below the 24 of film, so fast movement can look stepped. Interpolation to 30 fps helps.

Resolution tops out at 720p. Fine for social feeds and drafts, soft on a large screen.

Familiar weak spots. Hands, fast interactions between several people and readable text inside the video still fail often. Physics is plausible rather than exact: cloth and water move well, but a bouncing ball can misbehave.

Fewer knobs on Unicorn. The route does not expose a seed, a negative prompt or LoRA weights, so you cannot lock a result and vary only the wording, or load a style adapter.

Distilled, not identical. The fast version is a compressed model. For a hero shot where every detail matters, compare it with a less compressed render before committing.

Prompting Wan 2.2 Fast

The image already describes the scene, so the prompt should describe change: what moves, how the camera moves, and how the light behaves. Keep it to one subject action and one camera move for a five-second clip, and use film terms the model was trained on.

The fox lifts its head, ears twitching, and looks toward the camera as snowflakes drift down through the light. Slow push-in, shallow depth of field, soft morning glow.

Steam curls up from the coffee cup while the camera slowly orbits a quarter turn around the table, warm window light sliding across the wood.

The paper lanterns sway in a light breeze and their glow flickers; the camera cranes up from the street to reveal the rooftops at dusk.

A few habits help:

• Start from a clean, well-composed image at the target orientation. The output follows its framing, so a portrait image gives a portrait clip.
• Leave num_frames at 81 unless you need the extra length. Longer clips drift more.
• Draft at 480p, then render keepers at 720p with interpolate_output on.
• For transitions, set end_image and describe the journey between the two frames ("the camera pulls back as day turns to night").
• If motion looks too fast or too slow, change frames_per_second rather than rewriting the prompt.

Wan 2.2 Fast against other image-to-video models

At launch, Wan 2.2 Fast competes mostly with closed image-to-video models.

Hailuo 02 Standard from MiniMax makes six-second clips at 512p or 768p. It has a reputation for convincing physics in complex body motion, so pick it when the action itself is the hard part, and pick Wan 2.2 Fast for quick drafts, first-and-last-frame control and stylised inputs.

ByteDance's Seedance 1.0 Lite and Seedance 1.0 Pro render five- or ten-second clips and go up to 1080p. Choose Seedance when you need full HD; choose Wan 2.2 Fast when 720p is enough and you value an open base model.

Kling 2.1 Standard is Kuaishou's established mid-tier option. Runway Gen-4 Turbo suits fast, image-led iteration.

The honest summary: the closed models can look more polished on demanding shots, but none of them comes from a model you can also download and run yourself.

Choosing length, frame rate and resolution

Three settings shape the clip, and they interact. num_frames decides how much motion is generated: 81 frames is the default and the length the upstream documentation says works best, while 121 is the ceiling. frames_per_second only decides how fast those frames play back, so it is the quickest way to make motion feel slower or snappier without re-prompting. interpolate_output adds in-between frames after generation, so the clip keeps its running time but plays at 30 fps, which hides the stepped look of 16 fps on fast movement.

A sensible loop is to draft at 480p with the defaults, where Pruna measured about 28 seconds per clip, settle the prompt and the start image, then render the keeper at 720p with interpolation on. Pruna's figure for a 720p clip is about 80 seconds, so the final pass is still quick.

Getting started

Open Wan 2.2 Fast in the Playground, upload a start image, describe the motion and generate. For code, create a key on the API keys page and send a job to /v1/jobs with model, prompt and image, plus resolution, num_frames or end_image as needed. The documentation explains polling and webhooks, and the API tab on this page has ready-made examples.

Verdict

Wan 2.2 Fast is the open Wan 2.2 image-to-video model made quick enough to use casually. The mixture-of-experts design and cinematic training give it good camera work, stable first frames and a light touch with stylised art, and Pruna's distillation brings a five-second clip down to seconds. It is silent, short and capped at 720p, and hands and crowded action still trip it up. For animating stills in bulk, drafting shots before a premium render, or building first-to-last-frame transitions, it is very hard to beat on speed.