Buckets:

debrajsingha
/

dsp-repro-bundle

Files

xet

debrajsingha/dsp-repro-bundle / v0 /build_poster.py

debrajsingha

2 months ago

download

raw

11.6 kB

	import base64, pathlib

	OUT = pathlib.Path("outputs")
	def b64(p):
	return "data:image/png;base64," + base64.b64encode((OUT / p).read_bytes()).decode()

	fig_ridge = b64("fig_ridge.png")
	fig_anchor = b64("fig_anchor.png")
	fig_fid = b64("fig_fid.png")
	fig_cam = b64("conceptual_steering_gradcam.png")

	HTML = f"""<!DOCTYPE html>
	<html lang="en">
	<head>
	<meta charset="UTF-8">
	<title>Reproduction poster — DSP (Envisioning Beyond the Few)</title>
	<style>
	@page {{ size: A2 landscape; margin: 0; }}
	:root {{
	--accent:#1F6F84; --accent-deep:#144E5E; --accent-light:#E6F2F5; --gold:#C9812A;
	--text-primary:#182226; --text-muted:#6A7377; --bg-page:#F4F1EC; --bg-card:#FFFFFF;
	--border-soft:#D6D9D8; --u:1.6px;
	--font-sans:"Inter","Helvetica Neue",sans-serif; --font-serif:"Charter","Georgia",serif;
	}}
	* {{ box-sizing:border-box; margin:0; padding:0; }}
	p,li,figcaption,.body-text,.caption,.section-title {{ text-wrap:pretty; }}
	.title {{ text-wrap:balance; }}
	html,body {{ background:#2b2b2b; font-family:var(--font-serif); color:var(--text-primary); }}
	.poster {{
	width:calc(594 * var(--u)); height:calc(420 * var(--u));
	background:var(--bg-page); margin:20px auto;
	padding:calc(7 * var(--u)) calc(9 * var(--u));
	display:grid; grid-template-rows:auto 1fr auto auto; gap:calc(8 * var(--u));
	box-shadow:0 0 40px rgba(0,0,0,0.4); position:relative; overflow:hidden;
	}}
	.poster::before {{ content:""; position:absolute; top:0; left:0; right:0;
	height:calc(4 * var(--u)); background:linear-gradient(90deg,var(--accent-deep),var(--accent),var(--gold)); }}
	.header {{ display:grid; grid-template-columns:1fr auto; align-items:center; gap:calc(8 * var(--u));
	padding:calc(2 * var(--u)) calc(2 * var(--u)) calc(3 * var(--u)); border-bottom:calc(1 * var(--u)) solid var(--accent); }}
	.title-block {{ min-width:0; }}
	.title {{ font-family:var(--font-sans); font-weight:800; font-size:calc(15 * var(--u));
	line-height:1.06; color:var(--accent-deep); letter-spacing:-0.4px; }}
	.title .accent {{ color:var(--gold); }}
	.subtitle {{ font-family:var(--font-sans); font-size:calc(7.5 * var(--u)); color:var(--text-muted); margin-top:calc(1.5 * var(--u)); }}
	.authors {{ font-family:var(--font-sans); font-size:calc(6.8 * var(--u)); color:var(--accent); font-weight:600; margin-top:calc(1.5 * var(--u)); }}
	.qr-block {{ display:flex; flex-direction:column; align-items:center; gap:calc(1 * var(--u)); }}
	.qr-block img {{ width:calc(26 * var(--u)); height:calc(26 * var(--u)); border:calc(0.5 * var(--u)) solid var(--accent);
	border-radius:calc(1.5 * var(--u)); background:white; padding:calc(0.5 * var(--u)); }}
	.qr-label {{ font-family:var(--font-sans); font-size:calc(5.5 * var(--u)); color:var(--accent); font-weight:600; }}
	.body-grid {{ display:grid; grid-template-columns:1fr; min-height:0; }}
	.column {{ display:flex; flex-direction:column; min-height:0; }}
	.card {{ background:var(--bg-card); border-radius:calc(2 * var(--u)); padding:calc(5 * var(--u)) calc(6 * var(--u));
	border:calc(0.5 * var(--u)) solid var(--border-soft); box-shadow:0 calc(1 * var(--u)) calc(2 * var(--u)) rgba(0,0,0,0.04);
	border-left:calc(3 * var(--u)) solid var(--accent); flex:1; display:flex; flex-direction:column; }}
	.panels {{ display:grid; grid-template-columns:repeat(3,1fr); gap:calc(6 * var(--u)); flex:1; min-height:0; }}
	.panel {{ display:flex; flex-direction:column; }}
	.panel + .panel {{ border-left:calc(0.5 * var(--u)) solid var(--border-soft); padding-left:calc(6 * var(--u)); }}
	.section-title {{ font-family:var(--font-sans); font-weight:700; font-size:calc(9 * var(--u)); color:var(--accent-deep); margin-bottom:calc(1.5 * var(--u)); }}
	.section-title.sub {{ font-size:calc(8 * var(--u)); margin-top:calc(3 * var(--u)); }}
	.body-text {{ font-family:var(--font-serif); font-size:calc(6.1 * var(--u)); line-height:1.32; color:var(--text-primary); }}
	.body-text + .body-text {{ margin-top:calc(1.5 * var(--u)); }}
	.keyword {{ color:var(--accent); font-weight:700; }}
	.em {{ color:var(--gold); font-weight:700; }}
	.card ul {{ padding-left:calc(9 * var(--u)); margin-top:calc(1 * var(--u)); }}
	.card li {{ font-family:var(--font-serif); font-size:calc(6.0 * var(--u)); line-height:1.3; margin-top:calc(0.8 * var(--u)); }}
	.figure {{ margin-top:calc(1.6 * var(--u)); display:flex; flex-direction:column; align-items:center; }}
	.figure img {{ width:100%; display:block; border-radius:calc(1 * var(--u)); }}
	.figure .caption {{ font-family:var(--font-sans); font-size:calc(5.2 * var(--u)); color:var(--text-muted); margin-top:calc(1 * var(--u)); text-align:center; line-height:1.2; }}
	.callout {{ font-family:var(--font-sans); font-size:calc(6.2 * var(--u)); background:var(--accent-light); border-radius:calc(1.5 * var(--u));
	padding:calc(1.6 * var(--u)) calc(3 * var(--u)); margin-top:calc(1.8 * var(--u)); color:var(--accent-deep); }}
	.callout b {{ color:var(--gold); }}
	.footer-strip {{ display:grid; grid-template-columns:repeat(4,1fr); gap:calc(4 * var(--u)); background:var(--accent-deep);
	color:white; border-radius:calc(2 * var(--u)); padding:calc(3 * var(--u)) calc(5 * var(--u)); }}
	.hs-stat {{ display:flex; flex-direction:column; align-items:center; gap:calc(0.5 * var(--u)); }}
	.hs-stat .num {{ font-family:var(--font-sans); font-weight:800; font-size:calc(11 * var(--u)); color:var(--gold); white-space:nowrap; }}
	.hs-stat .lbl {{ font-family:var(--font-sans); font-size:calc(5.4 * var(--u)); color:var(--accent-light); text-align:center; }}
	.footer {{ display:grid; grid-template-columns:1fr auto; align-items:center; gap:calc(8 * var(--u));
	font-family:var(--font-sans); font-size:calc(5.6 * var(--u)); color:var(--text-muted); padding:calc(1 * var(--u)) calc(2 * var(--u)); }}
	.footer .right {{ color:var(--accent); font-weight:600; }}
	@media print {{
	:root {{ --u:1mm; }}
	html,body {{ background:white; margin:0; padding:0; width:100%; height:100%; overflow:hidden; }}
	.poster {{ margin:0; box-shadow:none; page-break-after:avoid; }}
	}}
	</style>
	</head>
	<body>
	<div class="poster" data-measure-role="poster">
	<header class="header" data-measure-role="header">
	<div class="title-block">
	<h1 class="title">Reproduction — Envisioning Beyond the Few: <span class="accent">Disentangled Semantics & Primitives</span> for Few-Shot Atypical Layout-to-Image Generation</h1>
	<div class="subtitle">Bao, Zhao, Wang, Li · ICML 2026 · OpenReview Jva4wVEySO · arXiv 2605.31266 · code: github.com/iCVTEAM/DSP</div>
	<div class="authors">Independent reproduction · verified on the official code + real DINOv2 / CLIP backbones · 1× T4 GPU Job, ≈$0.05</div>
	</div>
	<div class="qr-block">
	<img src="data:image/svg+xml;utf8,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'><rect width='100' height='100' fill='%23fff'/><g fill='%23222'><rect x='4' y='4' width='26' height='26'/><rect x='70' y='4' width='26' height='26'/><rect x='4' y='70' width='26' height='26'/><rect x='38' y='40' width='6' height='6'/><rect x='52' y='46' width='6' height='6'/><rect x='60' y='58' width='6' height='6'/><rect x='44' y='64' width='6' height='6'/></g><g fill='%23fff'><rect x='10' y='10' width='14' height='14'/><rect x='76' y='10' width='14' height='14'/><rect x='10' y='76' width='14' height='14'/></g></svg>" alt="QR to paper">
	<div class="qr-label">arXiv 2605.31266</div>
	</div>
	</header>

	<div class="body-grid" data-measure-role="body">
	<div class="column" data-measure-role="column">
	<div class="card" data-measure-role="card">
	<div class="panels">

	<div class="panel">
	<div class="section-title">The paper & what we tested</div>
	<p class="body-text">Existing layout-to-image models fragment under <span class="keyword">few-shot atypical</span> domains (aerial, underwater, low-light). DSP disentangles a category into <span class="keyword">stable semantics</span> and <span class="keyword">recomposable primitives</span> via three modules.</p>
	<div class="section-title sub">Claim 1 — three-module mechanism</div>
	<ul>
	<li><b>Semantic Anchoring</b>: DINOv2 ViT-L/14 features aggregated across the 5 exemplars into a per-category anchor.</li>
	<li><b>Primitive Imbuing</b>: visual prototypes fit by <b>closed-form ridge</b> T≈WP, refined by gradient (Alg. 1).</li>
	<li><b>Conceptual Steering</b>: text-driven <b>GradCAM</b> (CLIP) re-weights the denoising loss by GT-vs-gen saliency gap.</li>
	</ul>
	<div class="callout"><b>Verdict:</b> Claim 1 reproduced — all three modules present in the released code and behaving as described on the real backbones.</div>
	</div>

	<div class="panel">
	<div class="section-title">Primitive Imbuing: ridge is exact</div>
	<p class="body-text">Paper Eq. 7  <span class="keyword">&Wcirc; = T Pᵀ(P Pᵀ + λI)⁻¹</span>  is the exact minimiser of the Tikhonov objective Eq. 6. Reproduced from the repo's verbatim formula:</p>
	<div class="figure">
	<img src="{fig_ridge}" alt="ridge exactness and MSE reduction">
	<div class="caption">Closed-form matches the normal-equation solution to <b>2.7e-15</b> (fp64); Alg. 1 lowers reconstruction MSE on synthetic and <b>real DINOv2</b> tokens.</div>
	</div>
	<div class="callout">Closed-form vs autograd optimum: <b>3.6e-8</b> · gradient at optimum <b>9.4e-15</b> · real-feature recon MSE <b>−62.7%</b> vs K-Means.</div>
	</div>

	<div class="panel">
	<div class="section-title">Anchoring & Steering on real backbones</div>
	<div class="figure">
	<img src="{fig_anchor}" alt="semantic anchoring separation">
	<div class="caption">Per-category DINOv2 anchors: within-cat cos <b>0.997</b> vs cross-cat <b>0.895</b> (separation <b>0.102</b>).</div>
	</div>
	<div class="figure">
	<img src="{fig_cam}" alt="text-driven GradCAM">
	<div class="caption">Conceptual Steering: text-driven GradCAM on CLIP ViT-B/16 — foreground frac <b>0.72</b>.</div>
	</div>
	</div>

	</div>
	</div>
	</div>
	</div>

	<div class="footer-strip" data-measure-role="footer-strip">
	<div class="hs-stat"><div class="num">2.7e-15</div><div class="lbl">ridge closed-form error (fp64) — exact</div></div>
	<div class="hs-stat"><div class="num">−62.7%</div><div class="lbl">recon MSE vs K-Means (real DINOv2)</div></div>
	<div class="hs-stat"><div class="num">0.102</div><div class="lbl">anchor within−cross separation</div></div>
	<div class="hs-stat"><div class="num">FID −8.17</div><div class="lbl">Claim 2 target on DIOR (paper; not re-measured)</div></div>
	</div>

	<footer class="footer" data-measure-role="footer">
	<div>Claim 2 (Table-1 SOTA) documented, not re-measured: full SD-1.5 training on DIOR/RUOD/ExDark + Faster-RCNN eval is multi-GPU / multi-day with no released DSP checkpoints.</div>
	<div class="right">huggingface.co/datasets/debrajsingha/dsp-repro-artifacts</div>
	</footer>
	</div>
	</body>
	</html>"""

	pathlib.Path("poster.html").write_text(HTML, encoding="utf-8")
	print("wrote poster.html", len(HTML), "bytes")

Xet Storage Details

Size:: 11.6 kB
Xet hash:: 3983bfb0ec2502d262d46443660807f3eebf3c850e78a3d3c0d53d76e4489573

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.