36%
Lower background LPIPS
Up to 36.0% relative to full regeneration at early branch 9. This is background distance, not an editing success rate.
arXiv:2610.28003 · October 2026
Edit the path.
Keep the history.
Dancing Diffusion: Learning Continuous Programs
for Image Generation and Editing
Turn a generation request into continuous conditions for successive denoising stages. Save the executed state, rewrite what comes next, and resume.
01 / See the continuation
Original paper outputs · Figure 4
A different subject.
A familiar scene.
Target request A red truck drives down the opposite side of the road as a yellow bus and a small green car.
Final update 1423 · 6/12/12 schedule · branch 9 · seed 101.
View the full five-way paper comparison ↗Selected illustrations show scene retention alongside a requested replacement; they do not estimate success rates. Full regeneration starts from noise. The learned continuation resumes a saved generated state. Image-input controls in the full figure have different source access.
02 / The idea
Continuous condition space · frozen generator
Generation becomes
an editable program.
A learned compiler assigns a continuous condition block to each span of denoising steps. The generator executes the blocks over time. Editing replaces future conditions while reusing the state already reached.
01
Compile
Map the request to a sequence of continuous condition blocks.
02
Execute & save
Run the frozen generator and retain a genuine intermediate sampler state.
03
Rewrite & resume
Change the continuation and branch from the same executed history.
Watch the saved trajectory
From a changing prediction
to a finished image.
Every saved predicted-clean readout from the observer-guided example in Figure 3, in temporal order.
Controlled sphere study, separate from SANA. All 8 archived readouts at steps 0, 20, 40, 60, 80, 100, 140 and 199 are shown. These are not all 200 sampling steps. No interpolated frames; the example was selected post hoc.
Download the GIF ↗
03 / Measured in the paper
Read the full protocols ↗Preservation, quality
and reusable computation.
7.145
Generation FID
From native SANA's 7.349 on MJHQ-30K, with schedule 6/12/12 and the final checkpoint at update 1423.
22–23.4%
Less measured wall time
For four continuations with a shared prefix, versus recomputing it. Paired endpoints are identical. Native prompt switching can share this saving.
All final schedules
Keep the full comparison in view.
Same-protocol generation FID; lower is better. These are single-run, terminal-checkpoint results. Initialization histories differ across schedules.
Original paper table ↗
| Configuration | Final FID ↓ | Δ vs native |
|---|---|---|
| Native SANA | 7.349 | — |
| 3/9/18 | 7.315 | −0.034 |
| 6/12/12 | 7.145 | −0.204 |
| 9/12/9 | 7.375 | +0.026 |
Controlled study and limitations +
In a separate controlled image study, learned conditioning changes when content appears. This is distinct from the SANA experiment. Later branching can weaken edits; block interventions verify routing without establishing a training-induced semantic factorization. The method operates on saved generated states and is not a general arbitrary-photo editor.
04 / Work with the code
Research source release
Start with a
replay you can inspect.
The release includes the final trainer, continuous compiler, prefix capture/resume, evaluation implementations, configuration templates and CPU contract tests.
Available: source, tests, recorded tables and paper examples.
Not bundled: final weights, replay corpus and path-bound reproduction receipts.
CPU quickstart
git clone https://github.com/mistakes-glitch/declabeldiff.git
cd declabeldiff/dancing_diffusion
python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements-cpu.txt
python scripts/verify_release.py
python examples/prefix_edit_cpu.py
python -m pytestThe tiny CPU runtime checks execution and gradient contracts. It does not generate SANA images or reproduce the paper's numerical results.
BibTeX
@article{cui2026dancing,
title={Dancing Diffusion: Learning Continuous Programs
for Image Generation and Editing},
author={Cui, Enhe and Yang, Jinyue and
Yin, Bojian and Li, Guoqi},
journal={arXiv preprint arXiv:2610.28003},
year={2026}
}