“Wan2.1 supports 720p” is incomplete advice. The original Wan-Video repository distinguishes separate task checkpoints and model sizes, and the smaller 1.3B text-to-video model is documented for 480p rather than 720p. Anyone planning a local render farm needs to make the checkpoint choice before calculating GPU, storage or turnaround requirements.
Choose the task first
The repository lists T2V-14B, T2V-1.3B and two separate I2V-14B checkpoints for 480p and 720p. A text-to-video checkpoint accepts a prompt without a required starter image; an image-to-video checkpoint anchors the generation in a reference frame. Do not treat model parameter count as a universal quality score or promise identical behavior across tasks. The official download labels, command examples and file layouts are the right starting point for implementation, not an unversioned social-media preset.
Resolution and operations
Official model notes show T2V-14B supporting 480p and 720p, while T2V-1.3B supports 480p. The I2V variants are named by their resolution. Check those constraints before designing a final delivery size: increasing resolution in an editor may improve presentation but cannot add factual detail that the generator never produced. If you export to a larger frame, label that an upscaled delivery rather than changing the source model’s native generation specification.
Build a reproducible local run
Pin a repository commit, Python environment, model checkpoint identity and exact generation command. Log input prompt, input image hash when relevant, seed, task flag, frame dimensions and generated file checksum. The vendor’s README contains examples of model offloading and memory-saving options, but an example command is not a performance guarantee for every GPU. Start with a short baseline shot; if it fails, inspect environment errors and the task/weight match before blindly lowering quality settings.
Review and legal decisions
Evaluate temporal stability, foreground subject shape, camera motion and compliance with the creative brief. Compare output quality only after matching frame dimensions, clip duration and rendering hardware in a documented evaluation. Downloadable weights are not proof that every asset, model derivative and output can be used without restriction. Review the exact weight license and included notices, plus rights for reference imagery and any recognizable people. AI 11 owns detailed licensing and governance guidance.
When not to self-host
Hosted video generators remove much of the infrastructure work but trade away some control over versions, queueing and deployment. Self-hosting suits teams with GPU capacity, reproducibility needs and the engineering time to maintain checkpoints. A 480p 1.3B workflow may be appropriate for layout tests, while a 14B render or a conventional VFX plate can suit a demanding final shot. Record actual observed costs internally rather than inventing a cloud-versus-local winner.
Worked editorial scenario
A small production server must create twenty low-resolution storyboard loops every week, while a separate GPU workstation produces a handful of 720p finals. It may be efficient to scope a Wan2.1 1.3B 480p exploratory workflow and a 14B pipeline for selected final shots. But the team should first confirm hardware headroom, installed weights and what “final quality” means in the brief. Archive both environments with checkpoint hashes. If a scene needs a reference image, use the correct image-to-video checkpoint instead of trying to coax a text-only task into obeying an undocumented image input.
Sources & Further Reading
- Wan-AI and official project documentation — wan (first-party source; checked 2026-10-25)
Product availability, plans and rights should be rechecked before publication. No hands-on test, independent benchmark or rights-cleared third-party media is claimed in this article.