Animate a photo locally with Wan-Animate-2: what to check first

To transfer a wave into a product clip, choose a short source segment where the motion is clear. Avoid cuts and occlusion, and make sure the relevant hand and expression are visible. Start with a clearly visible wave rather than a full-body dance with several movements at once.

If person size or camera angle differs greatly between reference image and motion video, or the body is often cropped, the input mismatch can affect the result. Decide in advance which features to preserve—face, clothing, background—and which motion to transfer, such as a wave. Then review both appearance and movement, not just whether the character moved.

The official Wan-Animate-2 repository provides an inference¹ script, checkpoint-download instructions, and examples specifying reference-image and motion-video paths. Its download example uses the model identifier Wan-AI/Wan2.2-Animate-2-14B. If you use the repository path, follow its instructions and verify that a short input completes.

ComfyUI has a separate template connecting the image, motion video, description, and save nodes. It is not simply the official repository commands displayed as a graph, so align the ComfyUI version, custom nodes, and model files separately. Verify one path from input to saved output before comparing another runtime²; this keeps setup, input, and saving problems distinct.

The 14B parameter count alone does not determine VRAM³ needs. Reference images, video frames⁴, intermediate data, output size, and runtime batching use memory in addition to weights. The repository says its default 720p setup was configured on eight A800 GPUs and 480p was tested on two A800 GPUs. Do not treat these server examples as guarantees for one consumer GPU⁵.

Even if the ComfyUI template allows precision or cache changes, verify quality, speed, and memory with those settings. Simply adding memory capacities across GPUs or offloading⁶ does not imply the same runtime path or processing time. Look for a completed run on your hardware; without evidence, first test shorter motion and lower resolution rather than upgrading.

For the first test, run a short segment, save the final video file, and play it. Seeing an intermediate preview does not verify the saved file. Check whether face and clothing remain consistent with the reference, whether the wave follows the source motion, and whether the opening and ending cut off abruptly.

The ComfyUI template can also join 81-frame motion-transfer blocks into a longer video. A successful block does not establish continuity across the full sequence. If a join looks awkward, save and compare the blocks around the point where it changes instead of extending the whole sequence; adjust the reference image or description for appearance issues and the relevant segment for continuity issues.

For frequent revisions of short clips, check whether repeated runs reach a saved output; for a long sequence, account for joining blocks. Do not choose hardware from the time for one generation alone—include input selection, playback review, revisions and regeneration, and file saving.

The site’s video interface simulates differences in hardware waiting time; it is not a measured speed or quality evaluation of Wan-Animate-2. Do not compare results from another text-to-video model as if they were this model’s results. Increase resolution and duration one at a time on a supported path. Without a verified hardware example, do not buy based only on server-class A800 figures; test a shorter, lower-resolution path first.