---
license: mit
library_name: diffusers
pipeline_tag: image-to-video
tags:
  - worldcrafter
  - camera-control
  - autoregressive
  - context-memory
  - arxiv:2609.24984
---

# WorldCrafter-Context

**An autoregressive video baseline with standard historical-frame context memory
and UCPE camera control, without a memory encoder.**

WorldCrafter-Context is an intermediate training checkpoint released to help
researchers fine-tune the model and develop their own memory mechanisms, such as
point-cloud conditioning. It precedes the learned memory encoder used by
WorldCrafter-Base and WorldCrafter-Fast.

## Components

| Component | Source |
| --- | --- |
| Transformer | `TencentARC/WorldCrafter-Base`, `transformer/` |
| UCPE camera adapter | This repository, `camera_adapter.bin` (FP32, approximately 2.10 GB) |
| Text encoder, tokenizer, VAE and scheduler | `TencentARC/WorldCrafter-Fast` |
| Memory encoder | None; history is provided as video latents |

Shared components are referenced at fixed revisions in `inference_config.json`
and `model_index.json`; their weights are not duplicated here. `manifest.json`
records file hashes and sizes.

## Download

Install the Hugging Face CLI with `pip install huggingface_hub`, then run:

```bash
hf download Drexubery/WorldCrafter-Context --local-dir weights/WorldCrafter-Context
hf download TencentARC/WorldCrafter-Base \
  --revision d6b4731d4cec53f90b6afaee5a0f1b77d29979aa \
  --include "transformer/*" --local-dir weights/WorldCrafter-Base
hf download TencentARC/WorldCrafter-Fast \
  --revision d2ac12a654f8cafde4cb2e0ae7d5127ce2624939 \
  --include "text_encoder/*" "tokenizer/*" "vae/*" "scheduler/*" \
  --local-dir weights/WorldCrafter-Fast
```

These commands download only the shared components needed by Context. Existing
matching files can be reused. Keep the downloaded configuration and weight files
together.

## Construct the baseline

Combine Base's shared transformer, this repository's UCPE adapter and Fast's
text encoder, tokenizer, VAE and scheduler. Use historical video latents as context
without a memory encoder; do not load Base/Fast's additional LoRA, camera adapters
or RepEncoder. See the [training documentation](https://github.com/TencentARC/WorldCrafter/tree/main/training#worldcrafter-context-baseline).

`camera_adapter.bin` is the original FP32 PyTorch state dictionary, with 400
`blocks.<index>.cam_self_attn.*` tensors and no outer checkpoint wrapper.
This release is for custom integration; the current Base/Fast entrypoints and
standalone `DiffusionPipeline.from_pretrained` do not load Context automatically.

## Links

- [WorldCrafter code](https://github.com/TencentARC/WorldCrafter)
- [Training documentation](https://github.com/TencentARC/WorldCrafter/tree/main/training)
- [Paper](https://arxiv.org/abs/2609.24984)
- [Project website](https://drexubery.github.io/WorldCrafter/)