WorldFoundry
English中文

Warp-as-History

Method overview: Warp-as-History’s architecture from the paper.

Camera-controlled image-to-video: warp the still along a camera path, then refine. Dedicated env; official 33-frame / 16 fps demo is GPU-verified.

Edit this pageGitHubMarkdown

Warp-as-History (WAH) is a camera-controlled image-to-video method. The paper is Warp-as-History: Generalizable Camera-Controlled Video Generation from One Training Video (arXiv 2605.15182), with a project page at yyfz.github.io/warp-as-history and code at yyfz/Warp-as-History (Apache-2.0). The paper lists Shanghai Jiao Tong University and Shanghai AI Laboratory; this catalog card still has no separate institution field. Aliases wah, warp_as_history, and yyfz233/warp-as-history resolve here.

Official paper teaser: Warp-as-History samples from the paper.
Official paper teaser: Warp-as-History samples from the paper.
Method overview: Warp-as-History’s architecture from the paper.
Method overview: Warp-as-History’s architecture from the paper.

At inference time the still is geometrically warped along the requested camera path; those warps are treated as pseudo-history and fed through a frozen Helios-Distilled video model plus a task LoRA. Visible-token selection drops warp regions with no valid source. WorldFoundry vendors the official runtime under worldfoundry/synthesis/visual_generation/warp_as_history. Integration and runner status are verified: GPU parity passed with staged Helios-Distilled and LoRA weights, reproducing the official BMX / graffiti warp-video demo.

Build your run

Pipeline
WarpAsHistoryPipelineworldfoundry.pipelines.warp_as_history.pipeline_warp_as_history
TaskOfficial Demo Video official-demo-video
Environmentwarp-as-history
DeviceCUDA 12.1
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  warp-as-history \
  --pipeline.task-profile official-demo-video \
  --json

What it is for

Steer camera motion from one image. Required prompt and image; optional video; actions recorded as warp_video, camera_path. Artifact: warp-as-history.mp4 via WarpAsHistoryPipeline. The verified recipe is --pipeline.task-profile official-demo-video.

The official demo contract pins 33 frames, 16 fps, 384×640, seed 42. Score camera follow on CameraBench — that is the matching eval suite, not an ingested WorldFoundry rank.

Stage three external caches: BestWishYsh/Helios-Distilled (base), yyfz233/warp-as-history (visible_lora_state_step1000.safetensors), and yyfz233/Pi3X (model.safetensors) when you render warps from camera poses. Inject them with model_path / lora_path / pi3x_ckpt_path or WAH_HELIOS_DISTILLED_PATH / WAH_LORA_PATH / WAH_PI3X_CKPT_PATH. Pi3X is optional unless you start from poses instead of a ready warp video.

Runtime contract

Dedicated environment only: warp-as-history pins Python 3.10, CUDA 12.1, torch==2.5.0. Do not run this in the unified env. Backend stage is in_tree_official_cli. Camera paths far from the input viewpoint degrade the warp history and the clip.

Running in WorldFoundry

Launch worldfoundry-eval run warp-as-history --pipeline.task-profile official-demo-video after bash scripts/setup/model_env_install.sh --model warp-as-history. Per-run visual QA stays outside the catalog.