Wan2.2 Speech-to-Video vs Text/Image Models: Select the Right Checkpoint Before Rendering

A video generation model family is not one executable capability

Its generation code dispatches separate pipeline classes rather than hiding all work behind one universal object. This matters when selecting storage, interfaces and preprocessing: a video prototype made with T2V weights should not automatically be documented as supporting actor speech input. An installer that merely recognizes the label Wan2.2 cannot guarantee the required checkpoint is present.

Select by the input you actually possess

For a descriptive scene without a fixed starting image, start with the T2V route. For motion beginning from an approved frame, inspect I2V. TI2V is designed for text/image use cases with its own checkpoint, and S2V is a speech-conditioned output route. Those labels state broad modalities, not accuracy or quality rankings. If a producer needs lip movement tied to an approved voice recording, they should evaluate S2V specifically, while retaining voice-rights and transcript-review requirements from the dedicated audio team.

Resolution tables are checkpoint-specific

The Wan2.2 project lists T2V-A14B and I2V-A14B for 480p and 720p, TI2V-5B for 720p, and S2V-14B for 480p and 720p. It separately calls out 24 FPS for the TI2V-5B configuration. Those values should not be copied into a generic “Wan fps and resolution” statement for all future checkpoints or forked implementations. Store model family, exact weight identity, installation revision, task, output size and FPS under separate fields; blank unsupported parameters rather than filling them by analogy.

Speech-driven projects are sensitive to timing

Speech-to-video works with an audio signal, but a generation result may still have mouth timing and expression artifacts that only a human can judge. Split an approved narration into short segments that align with the desired shots, and preserve waveform timecodes. If the audio is revised, regenerate only the affected shots. Avoid routing copyrighted songs or imitated celebrity voices through an open checkpoint without authorization. S2V model availability does not grant rights to the speech content or public performance.

Reproducible local jobs

A production wrapper should record input hashes, checkpoint hash, task name, seed, configured size, frame count, sampler controls and output path. Separating configuration from an editorial prompt allows an engineer to rerun a failure or identify why two jobs differ. A scheduler can parallelize only after testing actual memory and disk constraints on the target machine. The project README is a capability reference, not evidence that the producer used this equipment or achieved any claimed rendering speed.

Worked example: historical guide performance

A museum needs a short fictional shopkeeper animation with already licensed English narration. Its design team provides a still portrait. The operator should check whether S2V can use the prepared audio and whether a different animation path is better for precise body gestures. Produce one short test, inspect syllable timing, then decide whether the segment needs a traditional cutaway instead. Keep the dialogue rights, illustrator license and approved script with the output metadata.

Costs and alternatives

A cloud video host may expose other Wan variants behind its own credit plans, which should not be confused with the official repository’s local checkpoints. Local compute, storage, postprocessing and editor time are still cost centers. Traditional puppet animation can be preferable for exact recurring faces; a managed speech-avatar vendor may have simpler deployment but more consent and subscription controls. The checkpoint choice should reflect the scene requirement rather than the model family name.

Sources & Further Reading

• Alibaba Wan team — Wan2.2 repository (first-party source; checked 2026-10-25)
• Alibaba Wan team — Wan2.2 generation entrypoint (first-party source; checked 2026-10-25)