Krea Realtime 14b
- Source
- krea.ai
- Author
- Krea editorial sitemap
- Date
Context
Krea Realtime 14B is designed to let a user redirect a video while it is being generated. It produces successive blocks of frames using earlier frames as , allowing new prompt instructions to influence the continuing stream. That creates an interactive workflow in which a creator can steer an unfolding scene. Technical account.
Generating from previous output also lets mistakes accumulate. Krea describes training the model on its own generated sequences so training more closely resembles actual use. It also identifies remaining limitations: complex camera movements are difficult, and short prompts tend to produce more static scenes. Detailed motion descriptions are therefore part of the model's stated operating conditions. Training approach and limitations.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
- KV cache — The memory a model keeps about text it has already read, so generating each new token doesn't require reprocessing the whole conversation.
- attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
Checking sign-in…
Loading comments…