引導式測驗 · 中等 等級
文字轉動畫和 AI 角色動畫測驗
測試運動擴散模型如何將文字轉換為骨骼動畫、為什麼會出現腳部滑動等偽像,以及視訊運動捕捉和重定向的工作原理。
問題 1 的 8
What does a text-to-motion model actually generate?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- What does a text-to-motion model actually generate?
- A rendered video of a character moving
- A sequence of skeletal poses over time
- A textured 3D mesh of a new character
- A script describing how an animator should move a rig
- Which model supplied the text embeddings that steer the original Motion Diffusion Model?
- 剪輯
- BERT only, fine-tuned on motion
- 語音辨識模型
- No text model; MDM reads raw characters
- HumanML3D, a key text-to-motion dataset, drew its clips largely from which source?
- ImageNet photographs
- YouTube dance videos only
- The AMASS motion capture collection
- Synthetic motions generated by a game engine
- Why does it matter that MDM predicts the clean motion rather than the added noise at each step?
- It makes the model run without a GPU
- 它消除了對任何訓練資料的需要
- It lets the model output video directly
- It allows geometric losses such as foot-contact and velocity penalties during training
- How do token-based systems like T2M-GPT and MoMask generate motion?
- By copying the nearest clip from a database
- By compressing motion into discrete codes with VQ-VAE-style quantization and generating those codes like words
- By simulating muscles with a physics engine
- By tracing keypoints from a reference video
- What is retargeting in AI animation pipelines?
- Changing the text prompt to get a different motion
- Removing jitter by averaging frames
- Mapping generated motion onto a character rig with different bones, proportions and rest pose
- Uploading motion to a cloud rendering service
- What is foot sliding?
- Feet skating along the ground when they should stay planted
- A deliberate moonwalk effect added by the model
- Feet disappearing from the rendered frame
- A retargeting error that swaps left and right legs
- Which sequence best describes how video-based AI motion capture works?
- Scan the actor with lidar, then animate by hand
- Ask a language model to describe the video, then run text-to-motion
- Attach markers to a suit and triangulate them with many cameras
- Detect 2D keypoints per frame, lift them to 3D, then smooth over time