Wan 2.2 Pose Control lets you transfer character poses with frame-by-frame precision

The Wan 2.2 Pose Control workflow solves a persistent problem in AI image generation: reliably transferring a character's pose while preserving their original design. The method, discovered by experimenting with Wan 2.2's I2V (image-to-video) model, exploits the innate temporal consistency of video models. By generating a sequence of 80 frames in First-Frame-Last-Frame mode, the model transitions from the original character to a pose reference image over four phases: static standing, movement into the target pose, morphing into the pose reference's character, and finally fully becoming that character. The target frame is captured at the exact moment the character has adopted the pose but before their design begins to morph.

The structured prompt is key: you segment the motion into time-stamped descriptions (e.g., "0s: girl with silver hair stands", "1s: she kneels and places hand on head") to ensure a smooth transition. In tests, Wan 2.2 achieved far better character consistency than the Flux.2 Klein character replacement workflow, which suffered from style bleeding and lost original proportions. Comparison with closed-source image-editing models showed that while Wan 2.2 preserves character design well, closed-source models still struggled with style fidelity (e.g., reproducing exact shading). The technique is currently limited to open-weight local models, making it accessible to developers running Wan 2.2 on their own hardware.