Paper Banana AI Research Figure Generator

A decision-tree selection workflow drawn top-down. Top — Starting Question (oval node): "Which retrieval architecture should I use?" Three sequential decision nodes (diamonds) below, each binary: Q1 — "Is the corpus < 10k documents?" - Yes -> "Use BM25 (sparse)" (leaf, green) - No -> continue to Q2 Q2 — "Do queries require semantic matching?" - No -> "Use BM25 (sparse)" (leaf, green) - Yes -> continue to Q3 Q3 — "Is latency budget < 100ms?" - Yes -> "Use ANN dense retrieval" (leaf, blue) - No -> "Use hybrid sparse + dense + re-rank" (leaf, purple) Three colored leaf nodes at the bottom showing the recommendations. Style: clean academic vector, white background, restrained palette (slate diamonds, green / blue / purple leaves), sans-serif. Branch labels ("Yes" / "No") on every edge. Optimised for poster and slide readability.

A 2x2 positioning matrix with two axes and four labeled quadrants. Axes: - X-axis (horizontal): "Data Efficiency" — left = low, right = high - Y-axis (vertical): "Compute Efficiency" — bottom = low, top = high Four quadrants: - Top-left: "Compute-Hungry, Data-Rich" — methods that need lots of compute and tolerate data scarcity (e.g., huge pretraining) - Top-right: "Sweet Spot" — high efficiency on both axes (e.g., distilled small models) - Bottom-left: "Avoid" — both inefficient (e.g., naive scratch training) - Bottom-right: "Data-Greedy, Cheap" — needs lots of data but cheap to run (e.g., kNN retrieval) Plot 6-8 example methods as labeled dots distributed across all four quadrants. Each dot has a 1-2 word label. Style: clean publication-style scatter, white background, slate axes with one accent color per quadrant, sans-serif. Quadrant headers in upper-corner of each quadrant. Suitable for survey papers and consultancy decks.

A horizontal mind-map comparing two approaches for vulnerability discovery in source code. Left side: "End-to-End LLM" Right side: "Architectural Decomposition" Each side has two color-coded branches: - Pros (dark green) - Cons (dark red) End-to-End LLM Pros: simple integration; fast prototyping; broad language coverage. Cons: hallucinated findings; no provenance; vulnerable to "semantic bypass". Architectural Decomposition Pros: traceable evidence chain; deterministic queries; ontology-grounded. Cons: heavier engineering; schema maintenance overhead. Layout: central vertical divider; method title at top of each side in a colored header bar; pros branches above, cons below. Use clean rounded labels, generous whitespace, sans-serif typography. Optimize for poster readability. White background.

A radar (spider) chart comparing three methods on five evaluation dimensions. Five axes (radiating from center, going clockwise from top): 1. Accuracy 2. Latency (inverted: higher = lower latency) 3. Memory Efficiency 4. Robustness 5. Interpretability Three methods plotted as colored polygons: - Method A (navy fill at 25% opacity, navy border) - Method B (teal fill at 25% opacity, teal border) - Method C (amber fill at 25% opacity, amber border) Numeric scores 0-1 along each axis, with grid rings at 0.25 / 0.50 / 0.75 / 1.00. Legend top-right with the three methods and their colors. Style: clean academic radar chart, white background, restrained palette, sans-serif labels. Optimise for legibility on a poster.

Create a clean comparison diagram titled "Shared Foundation, Divergent Choices" for an academic paper. Layout: - Top: one wide shared foundation block labeled "Shared pretraining foundation" with small icons for data, tokenizer, transformer layers, and optimization. - From the shared block, split into two parallel columns: "System A" and "System B". - Add four comparison rows: Attention pattern, Model scale, Training objective, Deployment context. - Each row should show matched but divergent choices, connected back to the shared foundation with subtle branching arrows. - Bottom: two outcome boxes labeled "Efficiency / controllability" and "Capability / generalization". Style: - Publication-ready vector-like diagram on white background. - Navy foundation block, teal System A accents, coral System B accents, neutral gray dividers. - Use concise labels only, maximum 6 words per cell. - Include small schematic icons but no decorative clutter. - Sans-serif typography, high contrast, suitable for a methods or related-work figure.

An educational infographic showing how a neural network learns, in five steps arranged around a circle. Steps (clockwise, starting from 12 o'clock): 1) Input — data point enters the network. Icon: small grid of pixels. 2) Forward pass — input flows through layers and produces a prediction. Icon: stacked rectangles with arrows. 3) Loss — prediction is compared to the truth label. Icon: balance scale. 4) Backward pass — gradients flow backwards through the network. Icon: backward arrow. 5) Update — weights are nudged to reduce the loss. Icon: small dial / slider. Center of the circle: a small "Repeat" label with a circular arrow. Style: - Friendly but professional palette: navy, teal, coral on white. - Each step is a rounded rectangle on the circle's perimeter, with the icon on top and a one-line caption (max 6 words) below. - Numbered badges (1–5) in the top-left of each step. - Light dashed circle connecting the steps to suggest the cycle. - Sans-serif, optimized for textbook-page readability.

A decision-tree selection workflow drawn top-down. Top — Starting Question (oval node): "Which retrieval architecture should I use?" Three sequential decision nodes (diamonds) below, each binary: Q1 — "Is the corpus < 10k documents?" - Yes -> "Use BM25 (sparse)" (leaf, green) - No -> continue to Q2 Q2 — "Do queries require semantic matching?" - No -> "Use BM25 (sparse)" (leaf, green) - Yes -> continue to Q3 Q3 — "Is latency budget < 100ms?" - Yes -> "Use ANN dense retrieval" (leaf, blue) - No -> "Use hybrid sparse + dense + re-rank" (leaf, purple) Three colored leaf nodes at the bottom showing the recommendations. Style: clean academic vector, white background, restrained palette (slate diamonds, green / blue / purple leaves), sans-serif. Branch labels ("Yes" / "No") on every edge. Optimised for poster and slide readability.

A 2x2 positioning matrix with two axes and four labeled quadrants. Axes: - X-axis (horizontal): "Data Efficiency" — left = low, right = high - Y-axis (vertical): "Compute Efficiency" — bottom = low, top = high Four quadrants: - Top-left: "Compute-Hungry, Data-Rich" — methods that need lots of compute and tolerate data scarcity (e.g., huge pretraining) - Top-right: "Sweet Spot" — high efficiency on both axes (e.g., distilled small models) - Bottom-left: "Avoid" — both inefficient (e.g., naive scratch training) - Bottom-right: "Data-Greedy, Cheap" — needs lots of data but cheap to run (e.g., kNN retrieval) Plot 6-8 example methods as labeled dots distributed across all four quadrants. Each dot has a 1-2 word label. Style: clean publication-style scatter, white background, slate axes with one accent color per quadrant, sans-serif. Quadrant headers in upper-corner of each quadrant. Suitable for survey papers and consultancy decks.

A horizontal mind-map comparing two approaches for vulnerability discovery in source code. Left side: "End-to-End LLM" Right side: "Architectural Decomposition" Each side has two color-coded branches: - Pros (dark green) - Cons (dark red) End-to-End LLM Pros: simple integration; fast prototyping; broad language coverage. Cons: hallucinated findings; no provenance; vulnerable to "semantic bypass". Architectural Decomposition Pros: traceable evidence chain; deterministic queries; ontology-grounded. Cons: heavier engineering; schema maintenance overhead. Layout: central vertical divider; method title at top of each side in a colored header bar; pros branches above, cons below. Use clean rounded labels, generous whitespace, sans-serif typography. Optimize for poster readability. White background.

A radar (spider) chart comparing three methods on five evaluation dimensions. Five axes (radiating from center, going clockwise from top): 1. Accuracy 2. Latency (inverted: higher = lower latency) 3. Memory Efficiency 4. Robustness 5. Interpretability Three methods plotted as colored polygons: - Method A (navy fill at 25% opacity, navy border) - Method B (teal fill at 25% opacity, teal border) - Method C (amber fill at 25% opacity, amber border) Numeric scores 0-1 along each axis, with grid rings at 0.25 / 0.50 / 0.75 / 1.00. Legend top-right with the three methods and their colors. Style: clean academic radar chart, white background, restrained palette, sans-serif labels. Optimise for legibility on a poster.

Create a clean comparison diagram titled "Shared Foundation, Divergent Choices" for an academic paper. Layout: - Top: one wide shared foundation block labeled "Shared pretraining foundation" with small icons for data, tokenizer, transformer layers, and optimization. - From the shared block, split into two parallel columns: "System A" and "System B". - Add four comparison rows: Attention pattern, Model scale, Training objective, Deployment context. - Each row should show matched but divergent choices, connected back to the shared foundation with subtle branching arrows. - Bottom: two outcome boxes labeled "Efficiency / controllability" and "Capability / generalization". Style: - Publication-ready vector-like diagram on white background. - Navy foundation block, teal System A accents, coral System B accents, neutral gray dividers. - Use concise labels only, maximum 6 words per cell. - Include small schematic icons but no decorative clutter. - Sans-serif typography, high contrast, suitable for a methods or related-work figure.

An educational infographic showing how a neural network learns, in five steps arranged around a circle. Steps (clockwise, starting from 12 o'clock): 1) Input — data point enters the network. Icon: small grid of pixels. 2) Forward pass — input flows through layers and produces a prediction. Icon: stacked rectangles with arrows. 3) Loss — prediction is compared to the truth label. Icon: balance scale. 4) Backward pass — gradients flow backwards through the network. Icon: backward arrow. 5) Update — weights are nudged to reduce the loss. Icon: small dial / slider. Center of the circle: a small "Repeat" label with a circular arrow. Style: - Friendly but professional palette: navy, teal, coral on white. - Each step is a rounded rectangle on the circle's perimeter, with the icon on top and a one-line caption (max 6 words) below. - Numbered badges (1–5) in the top-left of each step. - Light dashed circle connecting the steps to suggest the cycle. - Sans-serif, optimized for textbook-page readability.