Qualitative Results

By comparing with all baselines, we emphasize three key capabilities of our method that are accountable for the two key designs. Specifically, we observe good performance on generating instructional videos for actions with (i) large hand positional motion (without hallucination in the background); (ii) object state changing; (iii) subtle dexterous fingertip motion. Our proposed learnable automatic Region of Motion mask generation helps the model to focus on the accurate region for the action (``i'' and ``ii''), avoiding distraction from the cluttered background commonly seen in instructional videos. Our novel hand structure loss helps the model handle the actions with complex but essential fingertip motion (``iii''), which is rarely seen in previous non-instructional video generation benchmarks. For each sample, the first row displays the input context image and the results from different methods. In the second row, we first present the predicted Region of Motion mask from our Stage One model. Additionally, we show the motion flow below each generated video in the second row. Note that the motion flow is not part of the model’s outputs; it is shown solely to facilitate easier comparison among methods. Finally, we provide the Input Target Action Prompt (e.g., “Pick up and crack egg.”) and the Enriched Action Description (e.g., “The person uses their left hand to pick up an egg from the egg box and cracks it into a bowl.”).