A single developer's model card documents Zeroscope

The shot

In 2023, an independent developer publishing under the name cerspense released Zeroscope, a pair of open-weights text-to-video models distributed as free model cards on Hugging Face rather than as a company product. Repository metadata for the 576w card records its creation in mid-2023, in the same window as the first wave of major-lab video releases this note otherwise covers, but here published by one person rather than a funded lab.

What the documents show

The card states Zeroscope was 'fine-tuned from' the original weights of the Modelscope text-to-video model 'using 9,923 clips and 29,769 tagged frames,' and describes itself explicitly as 'a watermark-free Modelscope-based video model.' It generates at 576x320 resolution and 24 frames per second, and the card notes rendering 30 frames uses roughly 7.9GB of VRAM, low enough to run on a single consumer graphics card rather than a data-center cluster. The documentation is a model card written by the model's own author, not an independent benchmark study, so its claims about output quality are a builder's own account of a public experiment, not a peer-verified comparison against the commercial systems released the same year. The card also recommends the model as a preliminary step before upscaling with a companion zeroscope_v2_XL model using vid2vid techniques.

The workflow

The documented path runs in two stages: generate a base clip at 576x320 with the smaller model, then feed that clip into the XL upscaling model at around 1024x576 resolution, using denoise strength the card specifies as 'between 0.66 and 0.85,' reusing the original prompt. The card warns that 'lower resolutions or fewer frames could lead to suboptimal output,' meaning the documented settings are not merely suggestions but conditions the author states the pipeline depends on for usable results.

What the tool does not change

A model card that specifies VRAM use and denoise ranges says nothing about whether a given clip fits a production's shot list, matches surrounding footage, or clears any rights question about its training data; the card does not claim any of that. Because this is a single developer's free release rather than a company with a support channel, a production adopting it also takes on maintaining the pipeline itself, a cost the card does not mention but that its lack of any company backing implies.

• Is a claim about this model coming from its own author's card or from an outside evaluation?
• Does your hardware match the VRAM figure the card states for the frame count you need?
• Who maintains the pipeline once the individual developer moves on?

Zeroscope is evidence that, within months of the first lab-scale text-to-video releases, an independent developer could fine-tune and redistribute a comparable capability as a free, watermark-free download. That is a distribution story as much as a technical one, and the model card itself, written by the person who built it, is the only account of its behavior cited here.