debrajsingha/dsp-repro-bundle/v0 - Storage Bucket

Reproduction bundle — DSP (Envisioning Beyond the Few, ICML 2026)

Reproduction of "Envisioning Beyond the Few: Disentangled Semantics and Primitives for Few-Shot Atypical Layout-to-Image Generation" (Bao, Zhao, Wang, Li; ICML 2026; OpenReview Jva4wVEySO; arXiv 2605.31266). Official code: https://github.com/iCVTEAM/DSP

Logbook: https://huggingface.co/spaces/debrajsingha/repro-envisioning-beyond-the-few-disentangled-semantics-and-primitives-for-few-shot-atypical-lay Artifacts dataset: https://huggingface.co/datasets/debrajsingha/dsp-repro-artifacts

What is here

File	Purpose
ridge_verify.py	Claim 1a — proves the repo's closed-form ridge Ŵ=T Pᵀ(PPᵀ+λI)⁻¹ (Eq. 7) is the exact minimiser of the Tikhonov objective (Eq. 6), fp64; runs Alg. 1 alternating minimisation. CPU.
modules_verify.py	Claim 1b/c + real-feature 1a — runs all three modules on the real DINOv2 ViT-L/14 + CLIP ViT-B/16 backbones (Semantic Anchoring, Primitive Imbuing on real tokens, Conceptual Steering GradCAM). GPU.
make_poster_figs.py, build_poster.py, make_embed.py	build the reproduction poster (Chenruishuo/posterly A2 landscape) + poster_embed.html.
outputs/ridge_verify.json	ridge check results (exact to 2.7e-15).
outputs/modules_verify.json	real-backbone module results (T4).
outputs/conceptual_steering_gradcam.png	text-driven GradCAM heatmap.
outputs/fig_*.png, outputs/poster.png	poster figures + rendered poster.
poster.html, poster_embed.html	the poster and its self-contained embed.
paper_text.txt	extracted paper text (Eq. 6/7, Alg. 1, Table 1).

How to rerun

# 0. official code (provides DINOv2/CLIP/GradCAM implementations)
git clone https://github.com/iCVTEAM/DSP DSP_src
# blank the heavy package __init__ so the modules import without the full training stack:
echo "# narrow imports" > DSP_src/models/dsp/__init__.py

# 1. Claim 1a — closed-form ridge (CPU, seconds)
python ridge_verify.py            # -> outputs/ridge_verify.json ; "ALL RIDGE CHECKS PASSED"

# 2. Claim 1b/c — real backbones (GPU; downloads DINOv2 ViT-L + CLIP ViT-B/16 ~1.5 GB)
pip install torch torchvision einops opencv-python-headless ftfy ttach regex matplotlib
python modules_verify.py          # -> outputs/modules_verify.json + gradcam png ; prints RESULTS_JSON

The GPU run was executed as a Hugging Face Job on 1× T4: https://huggingface.co/jobs/debrajsingha/6a5bb304d216bd6f3a1fee76 (COMPLETED, 146 s).

Not reproduced (documented blocker)

Claim 2 (Table-1 SOTA numbers, incl. FID −8.17 vs CC-Diff on DIOR) requires full SD-1.5 base + few-shot training on DIOR/RUOD/ExDark, image generation, and Faster R-CNN / bootstrap-FID evaluation — multi-GPU / multi-day with no released DSP checkpoints. Cited as the target, not re-measured. See the logbook Claim 2 page.