WorldFoundry
English中文

FramePack

FramePack

Stanford constant-memory long image-to-video. Integration verified; runner and official long-video demo still pending. HF license cards are pending. Artifact framepack.mp4.

Edit this pageGitHubMarkdown

FramePack is Stanford's constant-memory long image-to-video packer (ControlNet authors): input-frame context is packed into a fixed-length window so generation cost stays roughly constant as the clip grows, including minute-scale clips on modest VRAM. WorldFoundry vendors a minimal infer-only runtime under worldfoundry/synthesis/visual_generation/framepack/framepack_runtime and exposes FramePackPipeline. Integration is Verified. The runner and official long-video demo are still pending.

Official paper teaser: Anti-drifting sampling and training methods. We present sampling approaches to generate frames in different temporal orders. The shadowed squares are the generated frames in each iteration, whereas the white squares.
Official paper teaser: Anti-drifting sampling and training methods. We present sampling approaches to generate frames in different temporal orders. The shadowed squares are the generated frames in each iteration, whereas the white squares.

Paper: arXiv 2504.12626. Project: lllyasviel.github.io/frame_pack_gitpage. Code: lllyasviel/FramePack. Public, non-gated weights: lllyasviel/FramePackI2V_HY and lllyasviel/FramePack_F1_I2V_HY_20250503. HF cardData did not expose a license — review before promotion. Catalog VRAM is unset.

Build your run

Pipeline
FramePackPipelineworldfoundry.pipelines.video_official.pipeline_official_video
TaskInteractive Video interactive-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  framepack \
  --pipeline.task-profile interactive-video \
  --json

What it is for

Pack long video as next-frame chunks from a start image.

Launch worldfoundry-eval run framepack.

Score on ChronoMagic-Bench — long-horizon generation suite matching FramePack's long-video focus.

Runtime contract

Required: prompt and image; artifacts: framepack.mp4.

KnobDefault
Duration5 s
FPS30
Steps25 per FramePack section
Latent window9
CFG / GS / RS1.0 / 10.0 / 0.0
Seed31337
TeaCacheon

Official in-tree runner defaults (the upstream UI quality recipe — not a GPU-validated long clip here):

Run in WorldFoundry

Status: Runner parity pending. Unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8). Stage lllyasviel/FramePackI2V_HY, lllyasviel/FramePack_F1_I2V_HY_20250503, and lllyasviel/FramePackI2V_HY among others locally — missing files fail at load, nothing is downloaded. Run worldfoundry-eval run framepack. Catalog VRAM is unset.