DanceOPD¶

Development line: project:danceopd · thread danceopd-development
Last event: 2026-06-26 · 1 dated since 2026-06-26 · Researched: 2026-09-06 · confidence: high

What it is¶

DanceOPD is a training framework for teams that already have compatible flow-matching checkpoints or LoRAs and need one student to combine text-to-image and image-editing behaviors.

• Hard-routes each sample to one frozen capability field.
• Queries a low-noise state from the current student rollout and matches velocity with MSE.
• Documents SD3.5/Diffusers and Z-Image/DiffSynth backends.

Development line¶

• 2026-06-26 — DanceOPD project website linked. On 2026-06-26, the DanceOPD GitHub Pages site went live. The site gave the project a public presence, but the underlying source text was not available.

What changed¶

• 2026-06-25 — arXiv v1 was submitted at 17:59:58 UTC, creating the canonical preprint record.
• 2026-06-26 — the v1 manuscript is dated June 26, and the public repository opened with the initial commit init.
• 2026-06-30 — Sync public release contents added 59 files with 3,208 additions and 76 deletions, including code, configs, data preparation, smoke scripts, and documentation.
• 2026-07-07 — arXiv v2 was submitted without a revision note, leaving its content changes unknown.
• 2026-08-15 — arXiv v3 was submitted. DanceOPD 0.2.0 added task-bound teacher routing, three modes across SD3.5-M and Z-Image, runnable public configs, paper templates, resume support, dual-LoRA support, and CFG-field options.

How to use this¶

1. Clone worldbench/DanceOPD, install the needed extra set, and run pip install -e ".[smoke]" plus bash scripts/bootstrap_smoke.sh. — https://github.com/worldbench/DanceOPD
2. Run the SD3.5 or Z-Image configuration dry run before loading real weights. Install DiffSynth-Studio before full Z-Image training. — https://github.com/worldbench/DanceOPD
3. Prepare a routed prompt or OmniEdit CSV. Edit routes require source_image, and style, background, or environment tasks map to global_edit unless overridden. — https://github.com/worldbench/DanceOPD
4. Configure each frozen teacher as a base checkpoint, full checkpoint, PEFT LoRA, or checkpoint-plus-LoRA. Keep the trainable student LoRA separate. — https://github.com/worldbench/DanceOPD
5. Use configs/public to exercise code or paper templates with compatible assets, then launch with Accelerate and resume from a saved step when needed. — https://github.com/worldbench/DanceOPD

Best practices¶

• Treat the smoke path as a code-health check, not evidence that paper metrics reproduce. — https://github.com/worldbench/DanceOPD
• For DanceOPD itself, retain the documented default: one hard-routed teacher, one low-noise query (K=1), and direct velocity MSE. — https://github.com/worldbench/DanceOPD
• Do not stack teacher LoRAs on the trainable student adapter. Merge each teacher into a clean frozen teacher module instead. — https://github.com/worldbench/DanceOPD
• Keep student_cfg_scale=1.0 when absorbing CFG, adjust the teacher scale, and tune inference CFG separately so effects do not compound. — https://github.com/worldbench/DanceOPD
• Keep downloadable public substitutes distinct from paper-aligned templates; the paper checkpoints remain unavailable. — https://github.com/worldbench/DanceOPD

Superseded by this¶

• 2026-06-26 — a project-page-only understanding was superseded by the public-code sync on 2026-06-30 and the 0.2.0 toolkit on 2026-08-15.
• 2026-08-15 — an undifferentiated reproduction path is obsolete: public substitute assets and paper-aligned templates are separate, and the paper checkpoints remain unreleased.

Still unknown¶

• The project page itself has no publication timestamp. The June 26 event relies on the manuscript's displayed date and repository history, while arXiv v1 was submitted on June 25 UTC.
• arXiv provides no change notes for v2 or v3, so manuscript edits remain unverified.
• No peer-reviewed conference or journal venue was verified; arXiv labels the work a Technical Report.
• No official hardware requirement, wall-clock training estimate, deterministic seed, hosted inference endpoint, or independent reproduction receipt was verified.

Sources¶

Agent brief¶

• Subject: project:danceopd, thread danceopd-development, 1 dated events 2026-06-26 → 2026-06-26.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.