Basic knowledge sharing of T2V PoC

[yuanxion]

T2V is planned to enable inferencing LLMs like Stable Diffusion on CPU/GPU, and training on Habana Gaudi/DG2, as well as improving the Generated Video quality, like more realistic frame, and coherency between frames.

[yuanxion]

Sharing of this PoC:

• Digital Image processing

• Stable Diffusion
  Model size: https://github.com/ThanatosShinji/onnx-tool#results-of-onnx-model-zoo-and-sota-models

[yuanxion]

Text2Video-Zero paper: https://arxiv.org/pdf/2303.13439.pdf
【AIGC-AI视频生成系列-文章1】Text2Video-Zero, https://zhuanlan.zhihu.com/p/626777733

Advantages: use text-2-image synthesis (Stable Diffusion) without any training or optimization
Development: Template-based -> GAN (attention) -> Transformer (multi-modal) -> Diffusion

Methods:

1. Stable Diffusion + zero shot (no more training or optimization)
2. Motion Dynamics in Latent Codes
3. Reprogramming Cross-Frame Attention
4. Background smoothing (Optional) with salient object detection
5. Conditional and Specialized Text-to-Video with ControlNet
6. Broadly applications for Diffusion-based (like Video Instruct-Pix2Pix) video generation and editing.

Thinking: Generate 3D foreground object, to improve the image quality?

[yuanxion]

ControlNet

• ControlNet 和 T2I-Adapter 有什么区别？
  https://zhuanlan.zhihu.com/p/619721909

ControlNet 在论文里提到，Canny Edge Detector 模型的训练用了 300 万张边缘-图像-标注对的语料，A100 80G 的 600 个 GPU 小时。Human Pose （人体姿态骨架）模型用了 8 万张 姿态-图像-标注 对的语料, A100 80G 的 400 个 GPU 时。 T2I-Adapter 的训练是在 4 块 Tesla 32G-V100 上只花了 2 天就完成，包括 3 种 condition，sketch（15 万张图片语料），Semantic segmentation map（16 万张）和 Keypose（15 万张）。 两者的差异：ControlNet 目前提供的预训模型，可用性完成度更高，支持更多种 condition detector （9 大类）。 T2I-Adapter ”在工程上设计和实现得更简洁和灵活，更容易集成和扩展” 此外，T2I-Adapter 支持一种以上的 condition model 引导，比如可以同时使用 sketch 和 segmentation map 作为输入条件，或 在一个蒙版区域 (也就是 inpaint ) 里使用 sketch 引导。

[yuanxion]

Stable Diffusion

• Stable Diffusion WebUI中文版
  https://ivonblog.com/posts/stable-diffusion-webui-manuals/zh-cn/

• stable diffusion模型分析 (tensorrt加速)
  https://zhuanlan.zhihu.com/p/593161199

• Training with Habana Gaudi
  https://github.com/HabanaAI/Model-References/tree/master/PyTorch/generative_models/stable-diffusion#training

[yuanxion]

Long Video generation

https://github.com/wuyongyi/awesome-video-generate

• phenaki
  https://phenaki.github.io/
  https://github.com/lucidrains/phenaki-pytorch

• ViViT论文阅读
  https://blog.csdn.net/qq_41533576/article/details/121678428