DisCO¶

Development line: project:disco · thread disco-development
Last event: 2023-07-11 · 3 dated since 2022-10-20 · Researched: 2026-09-04 · confidence: medium

What it is¶

DisCO covers three task-specific research implementations for image practitioners rather than one deployable tool.

• DISCO colorizes grayscale images through global color anchors.
• Portrait DisCO corrects close-up facial perspective through 3D-GAN inversion.
• Dance DisCo combines reference foreground, background, and target pose to generate human images or video.

Development line¶

• 2022-10-20 — DisCO project page and public demo appeared. Disentangled Image Colorization via Global Anchors established a SIGGRAPH Asia 2022 colorization workflow with code, checkpoints, and a hosted demo.
• 2023-02-26 — Portrait DisCO was presented through a project site. Portrait Distortion Correction with Perspective-Aware 3D GANs appeared as a separate project for correcting close-range portrait distortion. The repository records IJCV acceptance on 2024-01-03 and an inversion-code release on 2025-02-04, while the full system and Hugging Face demo remain TODO.
• 2023-07-11 — DisCo dance project was published with code and a live demo. Disentangled Control for Referring Human Dance Generation in Real World appeared as a separate pose-controlled dance-generation project. The repository records a temporal-module update on 2023-11-30, CVPR 2024 acceptance on 2024-04-08, and IDOL as a DisCo enhancement on 2024-07-15.

What changed¶

• 2022-10-20 — DISCO: Disentangled Image Colorization via Global Anchors established a SIGGRAPH Asia 2022 colorization workflow with code, checkpoints, and a hosted demo.
• 2023-02-26 — DisCO: Portrait Distortion Correction with Perspective-Aware 3D GANs appeared as a separate project for correcting close-range portrait distortion. The repository records IJCV acceptance on 2024-01-03 and an inversion-code release on 2025-02-04, while the full system and Hugging Face demo remain TODO.
• 2023-07-11 — DisCo: Disentangled Control for Referring Human Dance Generation in Real World appeared as a separate pose-controlled dance-generation project. The repository records a temporal-module update on 2023-11-30, CVPR 2024 acceptance on 2024-04-08, and IDOL as a DisCo enhancement on 2024-07-15.

These dates do not form one development line. The three projects have different authors, tasks, models, repositories, and release paths.

How to use this¶

As of 2023-07-11, check task-specific public project pages and validate capabilities through linked code or a live demo, rather than treating the name as one unchanged tool.

1. For grayscale colorization, clone the repository, create its documented Python 3.8 / PyTorch / CUDA environment, download a checkpoint into checkpoints, place images in data, then run the supplied inference script. — https://github.com/MenghanXia/DisentangledColorization
2. For portrait perspective correction, initialize EG3D and Deep3DFaceRecon, download the FFHQ checkpoint, create the supplied Conda environment, set input and model paths in example_configs/config.py, run preprocessing, then run the correction script. — https://github.com/lightChaserX/DisCO
3. For dance image editing, download the fine-tuning checkpoint and SD Image Variations dependency, set their paths in the supplied Jupyter notebook, and launch it on a local GPU. — https://github.com/Wangt-CN/DisCo

Best practices¶

• For DISCO colorization, retain the default 256×256 inference unless you have validated another setting: the authors mark original-resolution --no_resize output as unstable. Use absolute paths and fix --seed when comparing results. — https://github.com/MenghanXia/DisentangledColorization
• For Portrait DisCO, validate the EG3D and Deep3DFaceRecon dependency chain before processing images, and treat the released inversion code as narrower than a complete public product because the full system and hosted demo are still listed as TODO. — https://github.com/lightChaserX/DisCO
• For Dance DisCo, use the documented local notebook for an initial image workflow; for human-specific fine-tuning, prepare masks and skeletons with Grounded-SAM and OpenPose, then tune learning rate and U-Net unfreezing first. — https://github.com/Wangt-CN/DisCo

Superseded by this¶

• 2022-10-20 — using the hosted DISCO colorization Space as the default workflow is obsolete because its page currently reports a runtime error. Use the local repository path instead.
• 2023-02-26 — the earlier expectation of a future Portrait DisCO code release is superseded by the 2025-02-04 inversion-code release. Do not mistake it for a full-system or hosted-demo release.
• 2023-07-11 — treating the linked Gradio endpoint as a verified current Dance DisCo workflow is obsolete because it could not be fetched on 2026-09-04. Use the repository’s local workflow until a hosted endpoint is independently confirmed.
• 2023-07-11 — treating all occurrences of “DisCO” as versions of one project is obsolete. The three records are independent projects sharing a name.

Still unknown¶

• The disco key combines three independent projects. No primary source establishes a common maintainer, release train, API, model weight, or successor relationship.
• The current Dance DisCo Gradio endpoint could not be fetched during research; this does not prove permanent retirement.
• We did not execute any repository or download checkpoints, so compatibility with current CUDA, PyTorch, and dependency versions is unverified.

Sources¶

Agent brief¶

• Subject: project:disco, thread disco-development, 3 dated events 2022-10-20 → 2023-07-11.
• Practical note: As of 2023-07-11, practitioners following the DisCO line should look for task-specific public project pages and validate a claimed capability through its linked code or live demo, rather than treating the name alone as one unchanged tool.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.