Image-to-Video Generation

Most current image-to-video systems are latent video diffusion models. The input image is compressed into a latent by an autoencoder, and that latent is given to the denoising network at every step, often by concatenating it channel-wise with the noisy latents of each frame. Many models also inject a semantic embedding of the image, such as a CLIP image embedding, through cross-attention. Temporal attention layers let frames share information, so the model denoises all frames together into a coherent clip that starts from the given image. Stable Video Diffusion, released by Stability AI in November 2023, is a well-documented open example, with 14-frame and 25-frame variants. Commercial tools such as Runway, Kling, Luma Dream Machine and Google's Veo offer image-to-video, and open models such as Wan and HunyuanVideo have image-conditioned versions. Motion strength is controlled in several ways. Stable Video Diffusion takes a "motion bucket" value, learned from how much optical flow training clips contained, plus frame-rate conditioning. It also adds noise to the conditioning image. More of this noise augmentation lets the video depart further from the input, giving more motion but less fidelity. Commercial tools usually expose a motion slider, camera presets or a text prompt describing the motion. Typical artifacts include morphing (objects melting into each other), identity drift in faces after a few seconds, extra or warped limbs during fast movement, and texture swimming, where fine patterns seem to slide across surfaces. Other common problems are a flat "2.5D" parallax look, near-frozen output at low motion settings, and physics errors in liquids and cloth. A common misconception is that the model understands the 3D scene. It predicts plausible pixels. When a moving object reveals what was behind it, the model has no information about that area and invents it.