WorldFoundry
English中文

ZeroScope

cerspense's watermark-free 16:9 ModelScopeT2V fine-tune. Prompt-only 576×320 U-Net; Diffusers-matched WorldFoundry route; CC-BY-NC-4.0.

Edit this pageGitHubMarkdown

ZeroScope is a community text-to-video fine-tune of ModelScopeT2V by cerspense, aimed at watermark-free 16:9 clips. A prompt goes in; zeroscope.mp4 comes out. The in-tree runtime is prompt-only: image or video inputs raise. There is no camera path, no navigation action, and no 3D asset.

No official teaser or method figure is checked into this docs tree.

WorldFoundry executes the Hugging Face Diffusers U-Net through worldfoundry/synthesis/visual_generation/zeroscope/worldfoundry_runtime.py. For the selected configuration, output matched the official Diffusers route. Aliases zero-scope and zero_scope resolve here.

Two cards are recorded: cerspense/zeroscope_v2_576w (base 576×320) and cerspense/zeroscope_v2_XL (upscaler). The wired Studio/API profile stages the 576w checkpoint. Both licenses are CC-BY-NC-4.0.

Build your run

Pipeline
ZeroScopePipelineworldfoundry.pipelines.component_pipelines
TaskInteractive Video interactive-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  zeroscope \
  --pipeline.task-profile interactive-video \
  --json

What it is for

Smoke-test a legacy T2V pipe, or compare against the ModelScopeT2V parent — not produce modern DiT footage. Launch worldfoundry-eval run zeroscope with a required prompt. Runtime defaults: 8 frames, 576×320, 25 steps, guidance 9.0, 8 fps. Artifact: zeroscope.mp4.

Use FETV if you need a fine-grained suite from this U-Net era. For first-person world exploration use YUME. For a 3D splat use WorldGen.

Running in WorldFoundry

ZeroScopePipeline runs in worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8). Stage cerspense/zeroscope_v2_576w under WORLDFOUNDRY_HFD_ROOT. Review CC-BY-NC-4.0 before commercial use or automated benchmarking — the manifest records that policy review twice.

Minimal eval

The 576w route is runner-verified. Score generated clips on FETV with --fetv-metrics clip_score and read clip_score — that is the bounded official component. fetv_average appears only when every declared component is present.