Skip to content

Wan AI video guide with Wan 3 prompts and model checks

Plan Wan 3.0 video on Higgsfield with current model checks, original fox storyboards and editable prompts. Compare earlier Wan versions and inspect access terms.

Editorial cover artwork for Wan AI video guide with Wan 3 prompts and model checks
In this guide

Ready to create in Clout? Use the Wan video generator page and free shot planner for current studio controls and a copyable brief.

For six focused shot briefs and a timing plan, use the Wan 3 prompt guide. For input roles across models, start with the reference-to-video AI guide.

Clout now includes WAN 3.0 in its Image and Video Studio. The current configuration routes text, frame and reference requests separately. Choose the operation, upload compatible assets and inspect the displayed credit quote before rendering.

Higgsfield Looking for Wan 3 on Higgsfield? Open the official Wan 3 model entry. Our signed-out check showed Frames and References modes, optional Start and End frames, a prompt field, and model, duration and resolution controls. No generation was submitted.

Choose the Wan version before following a tutorial

Wan 3.0 expands the family beyond the earlier still-to-video workflow. The Alibaba Cloud generation guide documents native dialogue, music and effects, up to 30-second generation, and text, first-frame, first-and-last-frame, reference, editing and extension tasks. Those are documented capabilities, not a guarantee that a third-party interface exposes every combination.

The current hosted catalog lists wan3.0-video and the speed-optimized wan3.0-video-prime for 480P, 720P or 1080P output and 2–30-second generation. Prime is a named variant, not evidence of a speed improvement we measured. Inspect the selected operation rather than copying a setting from another version.

Starting materialTask to inspectFirst decision
Only a written ideaText to videoOne action or a timed sequence
One approved compositionFirst-frame image to videoWhat moves while identity stays fixed
Approved opening and endingFirst-and-last-frame modeWhether the motion can connect the endpoints
An existing clipEditing or extensionWhich part changes or continues

Alibaba's guide separates media combinations by task. First-and-last-frame mode accepts its endpoint types, while reference mode has its own combinations. Do not assume that adding an audio reference to an endpoint request is valid. Save the exact model, provider and selected mode beside your prompt.

Check the current Higgsfield offer without mixing versions

The October 13 Higgsfield release note describes seven days of Wan 3.0 and Prime access for eligible individual subscriptions from Pro upward, a shared queue, and a separate API discount ending October 31, 2026. This is a dated promotion, not permanent free access or an allowance verified on your account. Check the current offer and quote before committing a batch.

The older Wan landing page still identifies itself as WAN 2.5. Its trial wording should not be transferred to Wan 3.0. A model label, subscription offer, API discount and actual debit answer different questions. Keep all four in your production record.

Plan a short fox story before generating

Original three-panel planning illustration of an orange fox in a teal scarf noticing an oversized pinecone, reacting and standing farther back
Original proposed storyboard, not a Wan output or proof of video continuity

Our fox story has one visual joke and three readable beats. The wide view establishes the oversized pinecone, the portrait gives the fox a reaction, and the closing view restores the spatial relationship. The artwork is a reference sheet; it does not demonstrate that a video model preserved the scarf, scale or paw anatomy between shots.

For a first test, use the single-shot briefs later in this guide. For a longer experiment, the original Wan story-planning pack adds an editable 18-second sequence proposal, a reference sheet and a blank attempt log. It contains no generated video, prefilled success score or billing result. Eighteen seconds is our planned duration, not a measured export.

Original miniature forest story, one orange fox with a teal knitted scarf and one oversized brown pinecone. 0–6 seconds: locked wide eye-level view, the fox notices the stationary pinecone and pauses. 6–12 seconds: cut to a medium portrait, the same fox makes one restrained curious head tilt. 12–18 seconds: return to the wide view, the fox takes one small step back and settles. Match the daylight, scarf and pinecone scale. Quiet forest breeze and one soft footstep, no dialogue, music, text or extra characters.

Use this as a proposal to adapt to the selected task. For text-to-video, describe the three views. If you use references, assign their roles through the actual supported controls. Do not upload the whole contact sheet as a first frame unless you want the panel layout in the output. Prepare individual compositions for a still-driven sequence.

Compare an accepted longer take with three separate shorter takes only after both have been generated and reviewed. Record delivered duration, identity changes, sound and actual total charges. A single longer request can simplify assembly, but a weak reaction may be easier to replace when the shots are separate.

Keep earlier Wan workflows tied to their release

Wan 2.5 is a video-generation release in the Wan model family. Its hosted text-to-video and image-to-video preview models include synchronized audio. If your goal is a short creator scene, the useful choice is whether to invent the starting composition from text or animate a still you have already approved.

The following earlier-version walkthrough was checked against its named official sources on October 3, 2026; the Wan 3 sections above were added October 17. The fox story below is an original creative brief. We have not run a paid Wan generation, tested an account allowance or measured model quality. The artwork illustrates the brief rather than a Wan output.

Compare the hosted generation models

Alibaba Cloud's model catalog lists these hosted specifications. They describe named operations, not every local implementation or provider interface. Wan 2.5 is not the newest release in that catalog.

Hosted modelInput and soundDocumented output
wan2.5-t2v-previewText to video with audio sync480P, 720P or 1080P; 5 or 10 seconds; 30 fps MP4
wan2.5-i2v-previewImage to video with audio sync480P, 720P or 1080P; 5 or 10 seconds; 30 fps MP4
wan2.2-t2v-plusText to video without audio480P or 1080P; 5 seconds; 30 fps MP4
wan2.2-i2v-plusImage to video without audio480P or 1080P; 5 seconds; 30 fps MP4

For a Wan 2.5 vs Wan 2.2 decision, that audio difference is a useful starting point. Choose by the finished shot you need, then compare accepted exports. A specifications table cannot establish which version looks better, holds identity more reliably or costs less after retries.

Check free access before planning a campaign

Searching for Wan 2.5 free or Wan 2.5 image to video free leads to provider offers, rather than a universal model entitlement. The Higgsfield Wan landing page advertises two free generations and paid unlimited access in its FAQ. This is the public offer we observed, not an account-level allowance we tested.

Before spending a trial, inspect the selected model, operation, displayed charge and export conditions. Check whether the allowance covers the resolution you intend to publish and whether a retry consumes it. Save the actual downloaded file; a preview alone does not prove that the export meets your needs.

A search for Wan 2.5 free unlimited does not establish such a plan. If a provider uses the word unlimited, read the named model eligibility, expiry, queue rules and any restrictions shown for your account. A free trial and a paid unlimited offer are different access decisions. See the Higgsfield Unlimited guide for a structured access check.

  1. Write one target shot and its required delivery settings
  2. Confirm the exact Wan model and image or text operation
  3. Inspect the allowance and charge before submitting
  4. Record the request and inspect the exported file
  5. Count accepted shots, retries and finishing effort before expanding

Use image to video when the starting composition matters

For Wan 2.5 image to video, approve the important visual facts before motion. The fox example needs the orange coat, teal scarf, paws and forest edge clearly visible in the still. If the intended prop is absent, create a suitable composition first. Asking the video model to introduce a prop while performing a complex action makes the result harder to diagnose.

Choose text-to-video when you are exploring an original composition from a written brief. Name the subject, setting, framing, action and ending. For a recurring character, record what the first accepted clip actually looks like before treating a text description as a dependable identity baseline.

Higgsfield's public Wan page describes an upload, scenario and generation workflow. Your available controls may differ from an API integration. Preserve the provider label, selected version and settings with the prompt so an interface change does not silently become a different comparison.

Direct the sound as carefully as the image

A useful native-audio brief separates ambience, action sounds and speech. For the miniature forest, ask for quiet breeze and one soft scarf rustle. Begin without dialogue so you can hear whether the ambience fits the visual scene. Listen to the whole export with headphones before deciding the sound is usable.

If exact speech is required, specify a short line, who speaks and when. Review the actual words, pronunciation, voice and visible mouth timing. Generated audio support does not guarantee exact dialogue. Keep an alternative finishing workflow when a line must meet a strict script.

Higgsfield's speech and aspect-ratio help guide discusses Wan Speak and lipsync choices. Check that named workflow's current input controls. A speech-focused operation is not the same request as animating a still with ambient sound, and neither should be confused with motion-driven Wan Animate.

For social delivery, add captions from the final accepted speech. Use short phrases near the center, leave the face and joke visible, and inspect the actual mobile crop. Do not add a written caption to the motion prompt when exact typography and timing need an editable finishing step.

Download the video without confusing it with model weights

Wan 2.5 download can mean two different things: downloading a finished video or obtaining model checkpoints to run locally. For an exported video, use the provider's original download and inspect its dimensions, duration and sound. Avoid stretching a thumbnail or screen recording into the final delivery asset.

For is Wan 2.5 open source, our source check did not establish an official Wan 2.5 checkpoint release. The official Wan2.2 repository supplies a different release with its own license and model downloads. Its existence does not prove that hosted Wan 2.5 weights are available there.

Before a local setup, require a release-specific official checkpoint, matching inference code, a license and realistic hardware requirements. A similar filename or community download is insufficient evidence. Use the local Wan ComfyUI guide only for the release it documents.

For installation, use the dedicated Wan ComfyUI setup article. For a detailed still-to-motion brief, use the Wan image-to-video prompt guide. Keeping those jobs separate makes the next decision clearer.

Choose the task before downloading a model

The official Wan2.2 repository lists T2V-A14B for text-to-video, I2V-A14B for image-to-video, TI2V-5B for both, S2V-14B for speech-to-video and Animate-14B for character animation and replacement. These are task-specific model variants.

What you already haveTask to investigateInitial acceptance test
A written sceneText-to-videoDoes the scene and action match the brief?
An approved stillImage-to-videoDoes motion preserve the starting identity?
A speech recordingSpeech-to-videoDoes the visual performance match the speech?
A performance and character imageAnimate or replacementDoes the replacement move plausibly in the intended scene?

A filename with a similar number is not sufficient compatibility evidence. Check the task, model weights and workflow together. A generation template and an animation preprocessing pipeline should not be combined merely because both contain the Wan name.

Separate open models from hosted releases

A hosted Wan 2.5 option does not mean the Wan2.2 local repository provides that release. Alibaba Cloud's current text-to-video reference includes hosted Wan2.7 and distinguishes its protocol from older models. Hosted model access is different from a weights download.

For an API, match the account's region, endpoint and model. The Model Studio documentation describes asynchronous task creation and polling; resubmitting a job is not the same as checking its status. For a local run, check the exact variant's current hardware requirements before committing compute time.

The Wan2.2 repository's example for the large T2V model specifies substantial GPU memory. That example is not a universal requirement for every optimized workflow, but it is a reason to check the selected template carefully. Do not assume a laptop can run every model in the family.

Start with a small image-to-video scene

Use an original fictional fox character in a teal scarf standing at the edge of a miniature pine forest. Approve one still with the whole character visible and room for a simple gesture. Keep the scarf, ears and paws distinct from the background.

One continuous medium shot. The orange fox character raises one paw in a small wave, lowers it and tilts its head curiously. The camera remains steady at eye level. Keep the teal scarf, face and body proportions consistent. End with the character facing the camera.

This example isolates a gesture without adding a location change or complicated physical contact. If the paw deforms, simplify the action to a head tilt and check the source anatomy. If the scarf disappears, compare the starting image and the video before changing unrelated settings.

Save the accepted starting image, selected task, prompt and original export together. That record lets you distinguish a model change from a source-image change when a later clip behaves differently.

Use Animate when the performance is the input

For mode selection, reference preparation, the official ComfyUI path and downloadable shot briefs, use the dedicated Wan Animate workflow guide.

Character animation and replacement begin with a different production problem from still-image animation. You have a performance whose motion you want to use. Prepare the source with the documented pipeline and choose whether the result should animate the character or replace a performer within the source scene.

The separate character swap guide explains that distinction and the preprocessing assets involved. A face swap changes a narrower part of the visual identity; a full character replacement must account for the body, clothing, silhouette and interactions.

Begin with one visible performer, a stable camera and a short uncomplicated gesture. Review occlusion, foot contact and body proportions. If the original performer touches furniture or another person, the replacement must make those interactions believable, not merely follow the rough pose.

Build a short sequence from accepted shots

For a character story, plan an introduction, a small action and a reaction. Keep each shot's purpose written in one sentence. The fox might arrive at the forest edge, notice a comically oversized pinecone and look toward the viewer.

Generate the shots independently when that gives you clearer control. Before assembly, compare the face, scarf, scale and lighting across all three. Avoid using a failed first clip as the reference for the rest of the sequence; identity errors can spread through the whole story.

Add narration after the visual timing is stable, unless the chosen workflow specifically depends on speech as input. Then create captions from the final audio and place short phrases near the center without obscuring the action. Watch the completed sequence on a phone before publication.

Compare local and hosted workflows on the same job

DecisionLocal workflowHosted workflow
SetupCompatible weights, environment and hardwareAvailable model, region and account access
CostCompute, storage and setup effortGeneration charges and any plan limits
RepeatabilitySave workflow and model versionsSave provider, operation and request settings
DeliveryInspect encoded output and frame timingInspect original download and export conditions

Measure accepted work, including retries and finishing. Check whether a result has the required resolution, duration and usable audio. A free model download does not remove compute costs, and a convenient hosted interface does not prove it exposes every option in a research repository.

Download three shot briefs and a blank review sheet

Build a small sequence around one original fox and one visual joke: a pinecone almost as large as the character. These are untested editorial prompts, not claimed model results. Use a suitable approved still for each shot rather than asking one composition to transform into all three scenes.

Reveal shot: One continuous medium view of the original orange fox wearing a teal scarf beside a comically oversized pinecone at the miniature forest edge. The fox looks down at the stationary pinecone and pauses. Locked eye-level camera, soft daylight. Keep the scarf and pinecone shape stable. Quiet forest breeze, no speech, no music and no lettering.

Reaction shot: Begin from the approved medium portrait of the same fox and teal scarf. The fox glances toward the pinecone, then looks at camera with a small curious head tilt. Keep the paws still and the background stable. One continuous shot with a restrained push forward. Soft breeze only, no dialogue, no new objects and no captions.

Closing shot: Begin from the approved wide still showing the fox beside the oversized pinecone. The fox takes one small step back and pauses, keeping its whole body in frame. Locked camera, matching daylight and forest setting. Preserve the scarf and body proportions. One soft footstep with quiet ambience, no spoken line and no scene change.

Download the three Wan shot prompts and the blank motion and audio review sheet. Watch each full clip and inspect the action midpoint as well as the beginning and end. Record an actual observation and an accept, revise or reject decision for every shot.

For the reveal, inspect whether the pinecone remains stationary. For the reaction, check the eyes, scarf and intended expression. For the ending, watch foot contact and body proportions. Then listen separately for unwanted music, speech or sudden changes in ambience. A clip can pass the visual check while needing different sound.

When comparing providers, reuse the same source assets and shot briefs. Record offered settings and delivered settings separately. Include retries in your cost notes and avoid calling a single accepted clip a benchmark of the entire model family.

Create your recurring character in Clout

Build an original AI character in Clout and generate a coordinated set of photos and videos around that identity. Create your video character, approve the look and direct small scenes that belong to the same story.

Sources and testing

Independent Clout guide. Wan 3 entry on Higgsfield, October13 offer notes, Alibaba Cloud model catalog and Wan3 generation guide checked October17,2026. Public model entry loaded Frames and References modes, optional Start and End frames, a prompt field, model, duration and resolution controls while signed out. No paid Wan generation, account allowance, model-quality or speed benchmark tested. New fox storyboard and shot pack are original untested planning resources, not Wan outputs. Existing earlier-version sections remain explicitly scoped to their named models. Current provider controls and account quote take precedence.

Questions, answered

Frequently asked questions

What does Wan 3 add to the workflow

Alibaba Cloud documents up to 30-second generation, native audio and several reference, editing and extension modes for Wan3. Check the selected provider and operation; those capabilities do not prove that every input combination is valid or exposed in Higgsfield.

Is the fox storyboard a Wan video result

No. It is original planning artwork with an untested 18-second sequence brief. The pack includes no Wan-generated video, account receipt or measured model-quality result.

Is Wan free

Free access depends on the provider, selected version and account. The older Wan 2.5 landing page advertises a trial, while the October13 Wan 3 offer has separate eligibility and expiry conditions. Neither was verified as an allowance on your account.

Can Wan generate audio with video

The documented hosted Wan 2.5 text-to-video and image-to-video preview models include audio sync. Review the delivered speech and sound; native audio is not a guarantee of precise script delivery.

Can I download Wan model weights

Downloading an exported MP4 differs from downloading model checkpoints. Our check did not establish an official Wan 2.5 checkpoint release. The linked Wan2.2 repository documents another version.

Is Wan Animate the same as Wan image to video

No. Wan Animate uses a performance reference for character animation or replacement. Image-to-video begins from a still composition and motion direction. Follow the dedicated workflow for your selected task.

Have these Wan prompts been tested

No. They are original editorial shot briefs for a fictional fox story. The downloadable review sheet leaves observations and decisions blank for your own exports.

What is Wan Animate?

The official Wan2.2 repository lists Animate-14B for character animation and replacement. It uses a different workflow from ordinary text-to-video or image-to-video generation.

Can I use a Wan image-to-video model for every task?

No. Match the model variant and workflow to the task. Generation, speech and character replacement require different inputs and setup.

Does a hosted Wan release mean I can run it locally?

No. Verify an official weights release, compatible inference code, license and hardware requirements. Hosted API availability and local model availability are separate.