Polish model introduction and document causal video generation

README.md

@@ -1,22 +1,46 @@
 ---
 library_name: pytorch
 tags:
-- video-generation
-- pretraining
-- multi-view
+  - video-generation
+  - pretraining
+  - multi-view
+  - robotics
+  - openwam
+language:
+  - en
 ---
+
 # OpenWAM Pretraining
 
-OpenWAM video pretraining model weights, trained using single-view and multi view video data with text conditioning.
+[Project](https://openwam.stanford.edu/) | [Code](https://github.com/OpenWAM/OpenWAM) | [Pretraining guide](https://github.com/OpenWAM/OpenWAM/blob/main/docs/pretraining/index.md)
+
+**A visual foundation for robot world models.** OpenWAM Pretraining learns to
+predict how robot scenes evolve from observed video and text conditioning.
+Trained on single-view and multiview robot video, it provides a shared visual
+backbone for future prediction and downstream world-action modeling.
+
+The model learns visual dynamics without action supervision. Its weights can
+be used for video prediction or as an initialization for adapting OpenWAM to
+video-action policies and dynamics models.
 
-**Training objective of this revision: chunk-causal (autoregressive) video prediction.** Earlier revisions of this repository were trained with a bidirectional prefix/suffix objective; this revision continues from them with a teacher-forced chunk-causal objective:
+## Causal Video Generation
 
-- The first latent frame is a clean context frame; later latent frames form chunks of 4.
-- The model sees a noisy copy and a clean copy of the clip. A chunk being denoised attends to its own noisy frames and to the clean frames of earlier chunks only; clean frames attend to clean frames of their own and earlier chunks. No chunk attends to a later one.
-- Timesteps are sampled per frame. For half of the samples the clean copy is noised as an augmentation.
+The current release generates the future **one chunk at a time**. It starts
+from one observed latent frame, then predicts chunks of four latent frames,
+conditioning each chunk on the initial observation and preceding video.
+This autoregressive structure allows generation to progress without attending
+to later chunks.
 
-Use these weights chunk by chunk: denoise one chunk at a time while attending to the clean context and the previously generated chunks. Generating all future frames in one bidirectional pass is not the mode this revision was trained for and gives worse results than the earlier revisions; those remain available through the repository history.
+Training uses chunk-causal teacher forcing, with earlier clean video providing
+context for each noisy target chunk. Earlier repository revisions used a
+bidirectional prefix/suffix objective; this checkpoint continues that
+pretraining with the causal objective and should be used with chunk-by-chunk
+generation rather than a single bidirectional prediction of the entire future.
 
-This repository contains model weights only, with a portable reference configuration and checksums. Optimizer and scheduler state, datasets, and frontend assets are not included. The PyTorch payload contains `model_state_dict`. In the reference configuration the objective is selected by `training.use_teacher_forcing`, `training.chunk_size`, `training.window_size` and `training.noisy_video_condition_prob`.
+## Use with OpenWAM
 
-See the [OpenWAM pretraining source and documentation](https://github.com/DaivdYuan/OpenWAM-staging) for dataset processing and training instructions. Configure the external video frontend, text encoder, tokenizer, and local asset paths before use.
+The release includes model weights, a reference configuration, and checksums.
+The [pretraining guide](https://github.com/OpenWAM/OpenWAM/blob/main/docs/pretraining/index.md)
+covers data preparation, shared frontend assets, training, and video inference.
+See the [project page](https://openwam.stanford.edu/) for the broader approach
+to composing video prediction and robot control.