StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text

### Model/Pipeline/Scheduler description

Text-to-video diffusion models enable the generation of high-quality videos given text prompts, making it easy to create diverse and individual content. However, existing approaches mostly focus on short video generation (typically 16 or 24 frames), requiring hard cuts when naively extended to the case of long video synthesis. StreamingT2V, enables autoregressive generation of long videos of 80, 240, 600, 1200 or more frames with smooth transitions. The key components are:

- A ControlNet-like module which conditions the current generation on frames extracted from the previous chunk, using a cross-attention mechanism to integrate its features into the UNet's skip residual features.
- An IP-Adapter-like module which extracts high-level scene and object features from a fixed anchor frame in the first video chunk and is mixed into the prompt embedding features before executing spatial cross-attention.
- A SDEdit-based video refinement stage with randomized chunk sampling of overlapped frames per denoising timestep.

### Open source status

- [X] The model implementation is available.
- [X] The model weights are available (Only relevant if addition is not a scheduler).

### Provide useful links for the implementation

- Code: https://github.com/Picsart-AI-Research/StreamingT2V/-
- Weights: https://huggingface.co/PAIR/StreamingT2V
- Point of Contact Author: @hpoghos

Start by reviewing the linked StreamingT2V implementation and Hugging Face weights, then compare the model's components with the integration conventions in diffusers. Define the required pipeline entry points and validation coverage; done means the described long-video generation workflow is integrated and verified.

Vom Indexierungsmodell aus dem Issue-Text verfasst.