kpshinnik/repro-sigma-bundle

SigMa reproduction bundle

Reproduction of ICML 2026 paper σ: Sigmoid Modulation for Ultra High Resolution Diffusion (OpenReview 47JZSOkw5C; code github.com/bxuanz/SigMa).

This bundle verifies SigMa's training-free RoPE modulation mechanism — the algorithmic core both paper claims rest on — by running the paper's own released FluxPosEmbed positional-encoding module at the exact FLUX latent grids up to 16 megapixels. No model weights are needed: SigMa lives entirely in a weight-free RoPE layer.

Contents

• scripts/sigma_core.py — SigMa RoPE math copied verbatim from SigMa/flux/transformer_flux.py (only lifted out of the diffusers-heavy module to dodge an unrelated huggingface-hub version clash; no logic changed — line provenance is in the file header).
• scripts/repro_sigma_mechanism.py — the reproduction: drives the real FluxPosEmbed across resolutions 1024²→4096² and the 28-step flow schedule; writes outputs/*.csv, arrays.npz, summary.json.
• scripts/make_figures.py — renders the figures (outputs/fig_*.png, outputs/sigma_figure.html).
• outputs/ — all metrics (CSV/JSON), arrays, and figures.
• poster/ — the Chenruishuo/posterly poster (poster.html, self-contained poster_embed.html, poster_preview.png/pdf, GATE_REPORT.json).
• scripts/scaffold_icml_logbook.py, scripts/validate_icml_logbook.py — challenge tooling.

Rerun

pip install torch numpy matplotlib plotly
git clone https://github.com/bxuanz/SigMa            # source of the verbatim RoPE math
cp scripts/sigma_core.py scripts/repro_sigma_mechanism.py scripts/make_figures.py .
python repro_sigma_mechanism.py    # ~12 s on CPU -> outputs/
python make_figures.py             # -> outputs/fig_*.png, sigma_figure.html

What is NOT in scope

End-to-end 16-MP FLUX image generation and the SOTA FID/PickScore benchmarks are not regenerated: black-forest-labs/FLUX.1-Krea-dev (12B) is gated and Hugging Face Jobs returned 402 (no credits). The mechanism enabling those results is what this bundle verifies.