RepoPicks Radar

Updated every day around 23:30 UTC

Guides

Text2Video-Zero

Picsart-AI-Research/Text2Video-ZeroReviewed by AI

Checked by RepoPicks Radar on

What is Text2Video-Zero?

Text2Video-Zero makes short videos from a written prompt by driving an image model (Stable Diffusion) frame by frame, with no video training involved. It can also follow the poses or edges of an existing clip, or restyle one from an instruction such as "make it Van Gogh Starry Night". This is research code from a 2023 paper: Python 3.9, CUDA and an NVIDIA card with at least 12 GB of memory, or about 7 GB with compression turned up. The authors keep a free Hugging Face demo. The code has not changed since April 2023 and the method now also ships inside the Diffusers library.

Text2Video-Zero is free and can replace Runway, Pika, and Google Veo 3.

How easy is it to install Text2Video-Zero?

A · download and open · Web

  • Windows D
  • Linux D
  • Web A

The web is the easy way in at grade A: the authors keep a free Hugging Face Space (PAIR/Text2Video-Zero) where you type a prompt. Doing it yourself is D on Linux and Windows: "Install requirements using Python 3.9 and CUDA >= 11.6", create a virtualenv, `pip install -r requirements.txt` with 35 pinned packages from 2023, and have an NVIDIA card with 12 GB of memory before `python app.py` opens the page at 127.0.0.1:7860. macOS is not graded because the setup requires CUDA.

Text2Video-Zero on GitHub

Stars
4,247on Nov 10, 2026
Downloads
—all releases
Growth
—no gain in the last reading

Stars over time

4,2464,2474,248 Nov 6Nov 7Nov 8Nov 9Nov 10 4,2464,2474,248 Nov 6Nov 8Nov 10

Facts

Runs on
Windows, Linux, Web
You need to install
Python, NVIDIA graphics card
To build from source
Python, Graphics card, FFmpeg
Disk space
950 MB–4.3 GB · estimate
Accounts or keys
None
Cost
Free · No charge for the code or the models. The license is not a plain open-source one: it is the CreativeML Open RAIL-M, which adds rules about how you may use it.
License
NOASSERTION (Other)
Can replace
Runway, Pika, Veo 3
Last change on GitHub
Apr 14, 2023

Text2Video-Zero risk signals (10)

  • AttentionDownload links to file-sharing sites

    README links to https://www.dropbox.com/s/uv90mi2z598olsq/Text2Video-Zero.MP4?dl=0

    automatic check

  • AttentionDependencies with known vulnerabilities

    Pinned dependencies with known vulnerabilities (OSV): accelerate: GHSA-4j2p-28q2-5m79, PYSEC-2026-3804; basicsr: GHSA-86w8-vhw6-q9qq, PYSEC-2026-1215; diffusers: GHSA-7wx4-6vff-v64p, GHSA-98h9-4798-4q5v, PYSEC-2026-2446; gradio: GHSA-26jh-r8g2-6fpr, GHSA-279j-x4gx-hfrh, GHSA-34rf-p3r3-58x2; opencv_python: GHSA-qr4w-53vh-m672, PYSEC-2023-183; opencv-contrib-python: GHSA-cxjf-x6jp-p7mc, PYSEC-2023-181; pillow: GHSA-22hg-r758-rrw9, GHSA-24rc-8hj2-v98c, GHSA-254f-3936-288h; pytorch_lightning: GHSA-2vj5-px25-gjrp, GHSA-4cv3-v7pv-rfhf, GHSA-75m9-98v2-hjpm; torch: GHSA-3749-ghw9-m3mg, GHSA-53q9-r3pm-6pq6, GHSA-5pcm-hx3q-hm94; tqdm: GHSA-g7vv-2v7x-gj9p, PYSEC-2026-1976; transformers: GHSA-282v-666c-3fvg, GHSA-29pf-2h5f-8g72, GHSA-37mw-44qp-f5jm

    automatic check

  • Attentionresearch code untouched since april 2023

    The last commit is 2023-04-14 (1304 days before this check), there are no releases at all, and the recipe pins versions from that time: `torch==1.13.1`, `diffusers==0.14.0`, `gradio==3.23.0`, `transformers==4.26.0`, with environment.yaml asking for python=3.8.5 and cudatoolkit=11.3 while the README asks for Python 3.9. Promises such as "We plan to continue optimizing our code" were not followed by further commits.

    AI review

  • Attentionthe license limits how you may use it

    GitHub cannot identify the license (NOASSERTION) because the file is the "CreativeML Open RAIL-M dated March 28, 2023", which says it "added use-based restrictions not permitting the use of the Model in very specific scenarios" and that any derivative must carry "at minimum - the same use-based restrictions". This is not a plain permissive licence, and the README adds that the original Stable Diffusion code is under the same terms.

    AI review

  • Attentionone flag puts the page on the public internet

    README: "To access the app remotely, run this shell command: `python app.py --public_access`", followed by "For security information about public access we refer to the documentation of gradio". The page has no sign-in of its own, so the address it hands out is open to whoever has it.

    AI review

  • Attentionloads model files from community sites

    The code lists and loads models with `get_model_list()` so that "any base model for stable diffusion and any dreambooth model hosted on huggingface can now be loaded", and the README suggests "Dreambooth models can be obtained, for instance, from CIVITAI". Those are files uploaded by strangers that your Python process then loads, and the project offers no checksum or vetting.

    AI review

  • Infothe bundled styles copy commercial franchises

    The ready-made Dreambooth models are named "Anime DB", "Avatar DB", "GTA-5 DB" and "Arcane DB", triggered by the keywords `gtav style`, `arcane style` and `avatar style`. They exist to reproduce the look of films and games owned by other companies, and what you may publish from that depends on where you live.

    AI review

  • Infoneeds a large nvidia card

    News entry 03/28/2023: "We added a new low-memory setup. Minimum required GPU VRAM is currently **12 GB**", and 04/06/2023: with Token Merging "and chunk size set to `2`, our code can run with **less than 7 GB VRAM**", at the cost of image quality ("too high values will decrease the image quality").

    AI review

  • Infoclips are very short by default

    The documented call uses `"video_length": 8` frames at `fps = 4`, which is a two-second clip, and the published results are GIFs. A later news entry says generated videos "can have **arbitrary length**", but the defaults and examples stay at eight frames.

    AI review

  • Infoone feature is missing from the code

    README Ablation Study: "Adding `smooth_bg=True` activates background smoothing. However, our code does not include the salient object detector necessary to run that code." The repository also points to the same method inside the Diffusers library ("available in Diffusers, starting from version `0.15.0`"), which is better maintained than this code.

    AI review

Evidence is shown in its original language.

What Radar could not check about Text2Video-Zero

  • Whether the 2023 recipe still installs on a current graphics driver, Python or CUDA version
  • How good or steady the video looks on your own prompts, since the published examples were chosen by the authors
  • Real disk use, estimated by the checks at 950 MB to 4450 MB from a heuristic, once Stable Diffusion and Dreambooth files are downloaded
  • How long one clip takes on your card, which the README never states
  • What is inside a Dreambooth file taken from a community site before you load it

Similar projects

The RepoPicks app

Coming soon: install in 1 click with the RepoPicks app.

Before you download

See the release page on GitHub