stable-video-diffusion-img2vid-xt

This is a diffusion model developed by Stability AI for generating short video clips from a static image (image-to-video). The model creates videos up to 4 seconds long (25 frames at 576×1024 resolution), using the input image as a conditional frame. It was trained on large datasets and finetuned based on the previous version SVD Image-to-Video [14 frames], using an f8 decoder to improve the temporal coherence of frames. Training required ~200,000 hours of computation on GPU A100 80GB, with CO₂ emissions ~19,000 kg and energy consumption ~64,000 kWh.

Limitations:

• Short videos (≤25 frames), lack of full photorealism.
• Static frames or slow panning may occur.
• Poor quality when displaying text, faces, and people.
• The model’s autoencoding process smooths out details.

Usage conditions: Prohibited to create materials violating Stability AI’s policy (e.g., illegal, offensive, or misleading content). The model is not intended for generating factual or historical representations.

Announce Date: 20.11.2023
Parameters: 0B
Developer: Stability AI
License: Stable Video Diffusion Community

Use our pre-built public endpoints to test inference and explore stable-video-diffusion-img2vid-xt capabilities.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.