HANDI -- Qualitative Results on Baseline Comparison

By comparing with all baselines, we emphasize three key capabilities of our method that are accountable for the two key designs--Motion Area (MA) Generation and Hand Refinement Loss (HRL). Specifically, we observe good performance on generating Hand-Centric videos for actions require (i) intricate fingertip motion; (ii) large Motion Area; (iii) Object State Changing. Our proposed learnable automatic Motion Aera mask generation helps the model to focus on the accurate region for the action execution (``ii'' and ``iii''), avoiding adding distraction to background. Our novel Hand Refinement Loss helps the model handle the actions with critical details (``i''), which is under-explored in previousw ork. In this gallery, 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 Motion Area mask from our Stage 1 model. Then in the same row, we show the motion flow below generated video from each method. Note that the motion flow is not part of the model's outputs; it is shown solely to facilitate easier comparison among methods. Below the visuals, we provide the input Text Prompt for the target action (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.").