Use case: a real dance restaged in a palace by MiniMax H3 from a depth video, on one RTX 5090
A creator turned an 8.7-second dance clip into a grey depth video and had MiniMax H3 restage the routine with a new character in a palace, natively at 768×1376 on one RTX 5090. The write-up gives the full ComfyUI wiring, the settings and the mistakes that turned the palace into a white studio.
What it does
All five steps run in ComfyUI. Video Depth Anything turns the clip into a grey depth video (209 frames in about 11 seconds), Qwen-Image makes the character and places her in the palace in the pose of the first depth frame, and H3 Ref2VA gets the depth video as ref_videos and only that first frame as ref_images, pinned at frame 0 with MiniMaxH3AddGuide. NVIDIA's RTX frame generation then doubles it to 48 fps before an ffmpeg encode.
Results and limits
Knee bends, arm openings and a side turn line up in time, but arm angles are still off and the motion follows roughly, not frame for frame. A 768×1376 run takes about 12 minutes and peaked at 31.7 of 32.6 GB of VRAM; one seed went black after 3 seconds and a new seed fixed it. The author tested one clip with a fixed camera and one seed per condition.
Worth knowing
Put only the in-scene first frame in ref_images: a second, white-background character image turned part of the clip into a white studio. Don't mention the source dancer or her clothes anywhere in the prompt, not even the negative. Generate at 768 from the start, since upscaling a 480 render didn't fix the face, and wide trousers kept the knees readable where a long skirt hid them. MiniMax H3: Users in the EU, UK, US and South Korea need to request a license from MiniMax via the H3 License Request Form (platform.minimax.io/h3-license) before use.
Try it yourself
Picture 1 is [your setting] at frame 0. Video 1 gives the motion and timing. The same [setting] throughout, a fixed camera, visible [knee bends and arm openings].