ComfyUI’s official Wan2.2 tutorial names a 5B hybrid text/image-to-video model and separate 14B text-to-video and image-to-video models.
One family has several operationally different checkpoints
It also supplies first-and-last-frame examples. A local workflow cannot safely substitute these checkpoints merely because they share the Wan2.2 family label. Text conditioning, required encoders, noise models and memory demand can vary. Treat the workflow template and its linked model files as one reproducible unit.
Choose the smallest suitable experiment
For a first motion test, a hybrid model may reduce workflow complexity because it covers text and image inputs. For a product shot where the opening still must be preserved, begin with an image-to-video pipeline and assess whether geometry holds. A first-last-frame example is appropriate for a transition with two approved compositions, not for arbitrary continuous motion control. For complex real-world physics, no model-specific guide can guarantee valid interactions between hands, glass or cloth.
Model installation is part of the editorial specification
Before importing a downloaded workflow, compare its model names and folders with the official documentation. Pin the ComfyUI application revision, model checkpoints, precision, encoders and custom-node requirements. Keep a manifest of licenses and hashes for downloaded weights. A cloud session can differ from a local GPU in available memory and scheduling behavior; publishing a screenshot of a template does not establish that it ran on the user’s hardware.
A controlled three-shot test
Plan a short museum artifact sequence: a gentle turntable orbit, a macro engraving close-up and a transition from a still illustration into motion. Assign one workflow version to each shot and store all rejected attempts. Check that colors, artifact shape and labels remain correct through time. Where subtitles or names must be exact, apply them in an NLE after generation. Measure labor and compute cost per accepted second, including failed tasks and local hardware time.
Legal and performance constraints
Open checkpoints do not mean every output is free of other people’s rights. A model family may have separate weight licenses, dependencies and restrictions, so verify each upstream model card rather than borrowing another model’s commercial terms. Cloud usage is additionally subject to provider rules. No FPS, precise duration or cross-hardware speed claim is published here without a specific workflow setting and measured machine.
Relevant alternatives and trade-offs
ComfyUI’s LTX-2 family is another route for synchronized sound; hosted video services avoid local model management but introduce account and queue dependencies.