ECCV 2026 — visual design gallery

ECCV overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

28 of 28 oral records have a writing-focused manuscript study. Official program snapshot: 2026-10-02. Selected rendered pages were inspected for each study; this is not a proof audit or reproduction of experiments.

All official oral main PDFs in this snapshot were scanned. Separately linked supplementary PDFs were not retrieved for this edition; appendix lessons are limited to pointers visible in the main manuscripts. The absence of an attached appendix is not evidence that no supplement exists.

Inspect the complete inventory and source status · Official oral program

Read the display, then its design lesson. Tables and numeric charts below reconstruct explicitly attributed factual subsets. Diagrams are original teaching layouts, not screenshots or reproductions of the papers' architecture. Selected examples are not a prevalence ranking of visual styles.

Theory visualization: connect a named factor to an experiment

A second-order theory of texture for depth from focus · Second-order texture, Proposition 2, PDF p.5; Figure 8, p.12

Original planning layout. Explain the assumptions, variable meanings and which later panel tests each consequence. Any teaching schematic is original, not the paper's figure.

Observed in the source: The proposition labels texture, SNR and patch size in the equation; later aligned panels show bandwidth, texture crops, focus curves and recovered depth.

Prepare: Give every visually emphasized factor a name and unit, then vary the corresponding experimental quantity while reporting the controlled settings.

Do not infer: Do not present color-highlighted mathematics as evidence that assumptions hold universally.

Architecture diagram: distinguish supervision from inference

Chain-of-Visual-Thought: Teaching VLMs to See and Think Better with Continuous Visual Tokens · CoVT, Figure 3, PDF p.5

Original planning layout. Name frozen/trainable modules and distinguish the representations being learned from the signals that supervise them.

Observed in the source: The architecture separates expert supervision paths from VLM reasoning; its caption explicitly states dense decoding is optional at inference.

Prepare: Draw the input-to-output computation first, add training supervision as separately labeled paths, and identify optional outputs.

Do not infer: Do not imply training teachers or interpretable decoders must run during every inference call.

Inference intervention table: is the added modality actually used?

Make Geometry Matter for Spatial Reasoning · GeoSR, Table 5, PDF p.14

Setting	Full ↑	Without geometry ↑	Zero geometry ↑
Baseline (static)	50.2	49.9	49.7
GeoSR (static)	51.9	49.2	49.6
Baseline (dynamic)	62.8	63.1	63
GeoSR (dynamic)	66.1	63.2	65

Source-reported numerical subset. Scores copied from Table 5, higher is better; static and dynamic benchmark averages are different task aggregates and must not be pooled. Four rows and all three intervention columns are reproduced; no uncertainty or significance is inferred.

Observed in the source: The same models are evaluated with full, removed and zeroed geometry. The table keeps static and dynamic conditions separate, including a baseline row that improves after geometry removal.

Prepare: Separate training-time removal from inference-time intervention; state which model is fixed and what input changes.

Do not infer: A removal drop demonstrates reliance under that intervention, not an unrestricted causal proof of geometric reasoning.

Quality–time table: retain trade-offs and non-wins

NURBS Splatting: A Unified Differentiable Rendering Framework for Vector Graphics · NURBS Splatting, Table 2, PDF p.13

Variant	MSE ↓	SSIM ↑	LPIPS ↓	Time (s) ↓
Baseline	0.0025	0.9633	0.0576	590
Full method	0.0017	0.9803	0.0339	410
Without step annealing	0.0015	0.9839	0.0334	527
Without weights and knots	0.0018	0.9759	0.0348	354

Source-reported numerical subset. Source values for 100 Noto Emoji images at 480×480. MSE/LPIPS lower, SSIM higher, optimization time seconds lower. These are the four rows of Table 2, not a cross-dataset renderer ranking.

Observed in the source: Removing grid-step annealing improves quality but increases optimization time relative to the full model; removing rational weights and knots is faster but slightly lower quality.

Prepare: Put quality and runtime beside the same configuration names and preserve ablation rows even where the full method does not win.

Do not infer: Bold values are not statistical significance, and no single row is uniformly best across quality and speed.

Diagnostic curve: absolute performance versus normalized recovery

Understanding Geometric Representations in Self-Supervised Vision Transformers via Subspace Intervention · Subspace intervention, Figure 2, PDF p.8

Original planning layout. Define the task metric, rank, baseline floor and normalization. A schematic curve is illustrative and must not impersonate source numbers.

Observed in the source: The two panels share subspace rank but separate absolute geometric accuracy from a fraction of each model's full linear performance; random/residual controls provide a reference floor.

Prepare: Show the absolute score alongside any normalization and state the denominator. Use matched rank axes and distinguish model identity from experimental condition.

Do not infer: Near-complete normalized recovery for a weaker probe is not higher absolute task performance.

Hardware and qualitative evidence: show the measurement and difficult regions

Broadband Wide Field of View Imaging with Computational Mirrors · Computational Mirrors, Figure 3, PDF p.6; Figure 6, p.12

Original planning layout. Identify optical setup, wavelength/filter conditions, crop positions and acquisition budget; distinguish captured data from reconstructions.

Observed in the source: A ray diagram and annotated prototype explain acquisition; aligned reconstructed crops and on-axis/off-axis metrics expose where model differences matter.

Prepare: Use a hardware overview, matched regions/crops and metric subsets tied to those regions. State the number of acquired images.

Do not infer: Do not imply equal focal-stack budgets or single-shot operation when the source compares different acquisition policies.

Caption and visual checklist

State the display’s question or supported finding, define panels and symbols, keep conditions and units visible, and explain uncertainty. Label schematic curves as schematic. Separate training-only paths from inference. Retain baseline wins and preserve metric direction. View the final display at ordinary PDF size and in grayscale; do not rely on color alone.

Use the shared visual blueprints for additional architecture, proof, ablation and appendix-planning examples.