Wan 2.2 TI2V-5B is an option with available weights for generating local video sequences from text or an image. For marketing, a suitable use is a defined shot with clear inputs and acceptance criteria: a camera move, an animated background or a visual variation. Review the resulting clip and assemble it in an editable project.

Keep the complete variant name

The Wan family contains different models. TI2V-5B, A14B variants and the Wan 3.0 service have different configurations and distribution methods. Record the family, variant, version and implementation in the project brief. They determine which files and technical instructions apply.

The official Wan 2.2 repository declares Apache 2.0 for TI2V-5B. Retain component licences and input rights in the company’s records. A photograph found online does not become authorised advertising material by being submitted to a model; use your own photographs or assets carrying the necessary rights.

Two inputs, different control

  1. 1

    Text

    Describe the subject, framing and movement.

  2. 2

    Image

    Establish the starting appearance with a photograph.

  3. 3

    Sequence

    Check what the model retained and changed.

The diagram compares inputs. Inspect important details in every output.

Select an implementation for available memory

The native TI2V-5B example documents at least 24 GB VRAM with offloading and the text encoder on CPU. Official ComfyUI integration describes a 5B workflow with offloading for 8 GB VRAM. These are different software paths; one threshold does not automatically apply to the other or to A14B models.

Decide which path belongs in the implementation proposal. The technical record identifies the model, encoder, VAE, precision and memory settings. The text encoder prepares a representation of the prompt, while the VAE converts images and frames between visible form and the internal, or latent, representation used during generation. If a tutorial changes numerical representation or uses a community node, treat that as a separate variant with its own requirements and maintenance needs.

Person browsing a photo gallery on a laptop beside printed photographs
Michael Burrows

Prepare a photograph for the intended shot

The input image should show the object and its surroundings clearly. For a product, check shape, accessories, packaging and colour before use. An image with hidden features leaves more areas for the model to complete; retain real footage for essential details in a campaign.

Choose the image orientation for the intended delivery. The native example uses 720-class dimensions such as 1280 × 704 and 704 × 1280. These values belong to that documented workflow. The final export can use different dimensions in the editor, without changing the resolution at which the sequence was generated.

What belongs in the prompt and the editor
ElementIn the instructionsIn editing
SceneSubject, environment and lightingSequence order
MovementA defined action and camera moveCuts and pacing
Product identityA clear visual referenceReal footage for exact details
OfferNo invented prices in the imageApproved copy, price and conditions

Write instructions for one shot

Start with the subject, position and action, then specify the camera and lighting. An illustrative example is: “A white mug on a wooden table, camera slowly moving forward, side lighting, static background.” This demonstrates prompt structure; it is not a generation performed or an output validated for this guide.

Avoid asking one sequence to contain several locations, transformations and successive actions. Divide the script into shots and join them in the editor. If you use a negative-prompt field, keep it relevant to the defects being checked. A long list copied from another project makes the effect of changes harder to understand.

Three layers for a clear prompt

  1. 1

    Subject and framing

    What appears and how the scene is composed.

  2. 2

    Action and camera

    What moves and in which direction.

  3. 3

    Lighting and continuity

    The atmosphere and features to retain.

Retain the parameters for every variant

Archive the prompt, photograph, seed, dimensions, sequence length and workflow version. When comparing a different movement, retain the other settings. An accepted result should be identifiable without opening every file in the directory: a project name and variant number provide a practical organisation method.

Write selection criteria before generation: recognisable product, coherent movement, stable background and room for copy. Record rejection reasons, such as a changed mug handle or an additional object. Those observations help determine whether to change the photograph, the instructions or the approach to the shot.

Finish the advertisement in the video editor

The model supplies a sequence; the editor builds the campaign message. Add the logo, captions, offer and authorised sound there. Check the start and end, where a distortion may appear for only a few frames. Keep an editable project so a price or language can change without repeating the entire generation.

Calculate the budget per accepted clip, including operator time, rejected variants, processing and editing. This is a proposed accounting method for a business, not savings measured by WebHat. Size a future implementation for the required shot count and the campaign’s delivery schedule.

Three conditions for accepting a sequence

  1. 1

    Fidelity

    The product and important features match the reference.

  2. 2

    Continuity

    Objects and movement remain coherent over time.

  3. 3

    Usability

    The shot supports the message and fits the required edit.

Where WebHat consultancy fits

WebHat uses customised local AI and can help select a workflow and organise its operation. Custom applications with AI integration can connect inputs and approvals to the business catalogue.

Practical questions

Does Wan 2.2 run locally?

The selected variant can run with compatible weights and implementation on the hardware required by that workflow.

Is 8 GB VRAM sufficient?

ComfyUI documents that threshold for its TI2V-5B offloading workflow, not for the entire Wan family.

Does an image fix every detail?

The image conditions generation. Inspect shape, labels and continuity throughout the full sequence.

Can the offer change later?

Yes, when the offer is added as editable copy, separately from generated frames.