返回 ppt-master
image_prompts.json
1 {
2 "project": "attention_is_all_you_need",
3 "generated_at": "2026-05-24",
4 "deck_rendering": "blueprint",
5 "deck_palette": "cool-corporate",
6 "color_scheme": {
7 "primary": "#1A365D",
8 "secondary": "#F5F7FA",
9 "accent": "#3182CE"
10 },
11 "items": [
12 {
13 "filename": "cover_bg.png",
14 "purpose": "Cover background (Slide 01) — hero atmospheric blueprint of a generic Transformer-like module graph, calm left side reserved for SVG title overlay",
15 "page_role": "hero_page",
16 "text_policy": "none",
17 "aspect_ratio": "16:9",
18 "image_size": "2K",
19 "prompt": "Technical blueprint schematic style. Clean precise lines on a very subtle implied grid with deliberate geometric rigor — right angles, parallel rules, measured spacing. Elements are simplified to essential schematic forms — rounded rectangular module blocks, connector lines, anchor dots, callout markers. Color is restrained, near-monochrome with one semantic accent. A subtle grid background at 6% opacity reinforces the schematic feel. No textures, no shading, no painterly artifacts; engineering-precise, analytical, intentional. Color behavior is restrained-corporate: secondary off-white #F5F7FA carries about 65% of the canvas as calm breathing space; primary deep navy #1A365D forms the precise schematic linework (about 28% area) — outlines of module blocks, connector lines, anchor dots; accent blue #3182CE appears only as one highlighted module outline and a couple of arrowheads on the focal connector path, together under 7% of the area; no fourth color. Atmospheric blueprint backdrop with no dominant subject — a graph of six rounded rectangular module blocks arranged in two stacked horizontal rows of three (an encoder-stack row at the bottom and a decoder-stack row at the top), connected by precise straight lines with small triangular arrow heads. The blocks are deliberately small and clustered on the right side of the canvas, leaving the left third entirely calm as breathing space — that calm region must stay clear of any schematic element. One block on the upper row is highlighted by replacing its primary navy outline with an accent-blue outline, marking it as the focal self-attention module. Faint vertical and horizontal hairlines suggest a measurement grid only at 6% opacity beneath everything. Composed as a 1280×720 PPT 16:9 full-bleed cover background. Reads as the title plate of a systems-design briefing document. NO text, letters, numbers, signs, watermarks, or written symbols anywhere in the image. Color values (HEX codes like #1A365D) and color names are rendering guidance only — do NOT display HEX codes, color names, or palette labels as visible text anywhere in the image.",
20 "alt_text": "Abstract Transformer-like module graph rendered in technical blueprint style, calm left third reserved for title",
21 "status": "Generated"
22 },
23 {
24 "filename": "sequence_evolution.png",
25 "purpose": "Slide 03 hero — three schematic motifs side by side suggesting RNN chain / convolutional local field / global self-attention mesh, each motif carries its panel label inside the image",
26 "page_role": "hero_page",
27 "text_policy": "embedded",
28 "aspect_ratio": "16:9",
29 "image_size": "2K",
30 "prompt": "Technical blueprint schematic style. Clean precise lines on a very subtle implied grid. Elements are simplified to essential schematic forms — small filled circles for nodes, thin straight lines for connections, small triangular arrow heads on directed edges. Crisp uniform stroke around 1.5px, perfectly straight or precisely curved. No textures, no shading, no painterly artifacts. Color behavior is restrained-corporate: secondary off-white #F5F7FA carries about 62% of the canvas as calm breathing space; primary deep navy #1A365D carries the bulk of the linework (about 28% area, including the leftmost and middle motifs); accent blue #3182CE appears only in the rightmost motif (under 10% area), marking it as the focus. A subtle 6% opacity navy grid sits beneath everything. Custom composition — three equal-width schematic motifs sit side by side across the lower two-thirds of the canvas, separated by 2px hairline rules in border-gray; the upper third of the canvas is deliberately calm as breathing space for a page title overlay in SVG. Leftmost motif: a horizontal chain of five small filled circles flowing left to right, connected by single arrows between adjacent circles, suggesting recurrent step-by-step computation. Middle motif: three short parallel horizontal rows of four small circles each, each circle connected to the two circles diagonally below it in the next row by short straight lines, suggesting convolutional local receptive fields. Rightmost motif: five circles arranged in a regular pentagon, every circle connected to every other circle by thin straight lines forming a fully-connected mesh, suggesting global self-attention — this entire mesh is rendered in accent blue #3182CE while the first two motifs stay in primary navy. Above each of the three motifs sits one short panel label, rendered as clean monospace technical lettering (fixed-width, sharp uppercase, no serifs), small enough to read as a discreet figure-panel header — about 28-32px high relative to the canvas. The exact label text, left to right, is: above the leftmost motif the word RNN; above the middle motif the word CNN; above the rightmost motif the words SELF-ATTENTION. Labels above the first two motifs are in primary deep navy #1A365D; the SELF-ATTENTION label above the rightmost motif is in accent blue #3182CE matching its mesh. The three panel labels are the ONLY text in the image; no other letters, no axis labels, no captions, no signs, no watermarks anywhere else. Composed as a 1280×720 PPT 16:9 hero background; the upper third (above the panel labels) must remain calm for SVG title overlay. Color values (HEX codes like #1A365D) and color names are rendering guidance only — do NOT display HEX codes, color names, or palette labels as visible text anywhere in the image.",
31 "alt_text": "Three labeled schematic motifs (RNN / CNN / SELF-ATTENTION) comparing recurrent chain, convolutional local fields, and global self-attention mesh",
32 "status": "Generated"
33 },
34 {
35 "filename": "rnn_bottleneck.png",
36 "purpose": "Slide 04 right-third illustration — sequential RNN computation bottleneck (vertical chain of hidden states with an hourglass cue)",
37 "type": "flowchart",
38 "page_role": "local",
39 "text_policy": "none",
40 "aspect_ratio": "3:4",
41 "image_size": "1K",
42 "prompt": "Technical blueprint schematic style. Clean precise lines on a very subtle implied grid, with deliberate geometric rigor — right angles, parallel rules, measured spacing. Elements are simplified to schematic forms — small rounded rectangles, thin straight arrows, an hourglass marker. Crisp uniform stroke around 1.5px. No textures, no shading, no painterly artifacts; engineering-precise, restrained. Color behavior is restrained-corporate: secondary off-white #F5F7FA fills about 60% of the block as background breathing; primary deep navy #1A365D forms the chain of hidden-state boxes and the connecting arrows (about 32% area); accent blue #3182CE is used only for the hourglass marker and one small focus highlight (under 8% area); no fourth color. A subtle 6% opacity navy grid sits beneath everything. Flowchart-style local composition: a vertical chain of six small identical rounded rectangles stacked one above the other, each connected to the next by a thin downward-pointing arrow with a small triangular arrow head; each rectangle represents a hidden state at one time step. To the right of the chain, a small hourglass marker in accent blue floats alongside the middle of the chain, suggesting sequential time cost. To the left of the chain, two faint dashed parallel vertical lines hint at the channels that cannot be parallelized — they are dashed, in primary navy at 50% opacity. The entire schematic is centered inside the block with about 14% inner padding on all sides. Composed as a 480×600 vertical local block for placement inside an SVG page region; do NOT reserve calm overlay space inside the image — the SVG layer surrounds this block, not on top of it. NO text, letters, numbers, signs, watermarks, or written symbols anywhere in the image. Color values (HEX codes like #1A365D) and color names are rendering guidance only — do NOT display HEX codes, color names, or palette labels as visible text anywhere in the image.",
43 "alt_text": "Vertical chain of six hidden-state boxes with downward arrows and a small hourglass marker, in blueprint schematic style",
44 "status": "Generated"
45 },
46 {
47 "filename": "sublayer_block.png",
48 "purpose": "Slide 10 left-third illustration — encoder/decoder sub-layer composition: outer frame, two inner sub-layers, residual arcs, LayerNorm stripes; modules carry conventional paper-figure labels",
49 "type": "framework",
50 "page_role": "local",
51 "text_policy": "embedded",
52 "aspect_ratio": "3:4",
53 "image_size": "1K",
54 "prompt": "Technical blueprint schematic style. Clean precise lines on a very subtle implied grid, with deliberate geometric rigor — right angles, parallel rules, measured spacing. Elements are simplified to schematic forms — rounded rectangles, thin arcs, small horizontal stripes. Crisp uniform stroke around 1.5px. No textures, no shading, no painterly artifacts; engineering-precise. Color behavior is restrained-corporate: secondary off-white #F5F7FA fills about 60% of the block as background breathing space; primary deep navy #1A365D draws the outer frame, the inner sub-layer rectangles, and the stripes (about 30% area); accent blue #3182CE is used only on the two residual arcs to highlight them (under 8% area); no fourth color. A subtle 6% opacity navy grid sits beneath everything. Framework-style local composition: a single large outer rounded rectangle frame represents one encoder/decoder layer block. Inside that frame, top-to-bottom, two inner stacked rounded rectangles — the upper one slightly taller, the lower one slightly shorter; they are separated by a small gap. Around each inner sub-layer, a thin curved arc in accent blue starts above the sub-layer, sweeps around its right side, and rejoins below it — these two arcs are the residual skip connections, the only accent-blue elements in the image. Below each inner sub-layer, a thin short horizontal stripe in primary navy suggests layer normalization. Each inner sub-layer rectangle carries one short module label rendered as clean monospace technical lettering (fixed-width, sharp letterforms, no serifs), centered inside the rectangle and sized to read as a discreet figure label — about 18-22px relative to the block. The exact label inside the upper (taller) sub-layer is the words Self-Attention; the exact label inside the lower (shorter) sub-layer is the words Feed Forward. Below each LayerNorm stripe sits one tiny annotation in the same monospace lettering, smaller (about 12-14px), in primary navy, reading exactly Add & Norm. These four short labels (Self-Attention, Feed Forward, Add & Norm, Add & Norm) are the ONLY text in the image — no other letters, no numbers, no signs, no watermarks anywhere else. The entire schematic is centered inside the block with about 14% inner padding on all sides. Composed as a 480×600 vertical local block for placement inside an SVG page region; do NOT reserve calm overlay space inside the image — the SVG layer surrounds this block, not on top of it. Color values (HEX codes like #1A365D) and color names are rendering guidance only — do NOT display HEX codes, color names, or palette labels as visible text anywhere in the image.",
55 "alt_text": "Encoder layer frame with two stacked labeled sub-layers (Self-Attention, Feed Forward), accent residual arcs, and Add & Norm stripes",
56 "status": "Generated"
57 },
58 {
59 "filename": "positional_encoding.png",
60 "purpose": "Slide 11 hero — sinusoidal positional encoding visualization (two superimposed waveforms at different frequencies plus a discrete position axis below); textbook curve labels and axis label inside the image",
61 "page_role": "hero_page",
62 "text_policy": "embedded",
63 "aspect_ratio": "16:9",
64 "image_size": "2K",
65 "prompt": "Technical blueprint schematic style. Clean precise lines on a very subtle implied grid. Curves are precisely drawn smooth sine and cosine waveforms with uniform stroke around 1.5px. No textures, no shading, no painterly artifacts; engineering-precise, reads as a technical figure from a textbook chapter on Fourier features. Color behavior is restrained-corporate: secondary off-white #F5F7FA carries about 58% of the canvas as breathing field; primary deep navy #1A365D draws the slower (longer wavelength, gentle undulation) sine curve and the tick marks (about 30% area); accent blue #3182CE draws the faster (shorter wavelength, tighter undulation) cosine curve and three thin vertical guide-lines that drop from the waveforms down to the position axis (under 12% area); no fourth color. A subtle 6% opacity navy grid sits beneath everything. Custom composition — the left third of the canvas is deliberately calm as breathing space reserved for a page title overlay in SVG. Across the right two-thirds of the canvas, two superimposed sinusoidal waveforms run horizontally: the primary-navy sine curve has a long wavelength spanning roughly two full cycles across the available width with gentle undulation; the accent-blue cosine curve runs at higher frequency completing roughly five full cycles across the same width with tighter undulation; the two curves visibly cross several times. Below the waveforms, a horizontal row of ten small evenly-spaced tick marks in primary navy marks discrete token positions on a position axis represented by a thin horizontal line. Three thin vertical accent-blue guide lines drop from selected points on the waveforms straight down to corresponding ticks on the position axis, suggesting that each token position carries a unique pair of sine and cosine values. Curve labels and one axis label are rendered inside the image as clean monospace technical lettering (fixed-width, sharp letterforms, no serifs, about 22-26px relative to canvas). The exact text is, in three places only: near the right end of the navy sine curve, the lowercase word sin in primary navy #1A365D; near the right end of the accent-blue cosine curve, the lowercase word cos in accent blue #3182CE; just below the right end of the position-axis line, the lowercase phrase position → in primary navy. These three short labels (sin, cos, position →) are the ONLY text in the image — no other letters, no numbers, no axis tick numbers, no signs, no watermarks anywhere else. Composed as a 1280×720 PPT 16:9 hero background; the left third must remain calm for SVG title overlay. Color values (HEX codes like #1A365D) and color names are rendering guidance only — do NOT display HEX codes, color names, or palette labels as visible text anywhere in the image.",
66 "alt_text": "Sine and cosine sinusoidal waveforms with sin/cos curve labels and a position arrow axis label, blueprint schematic",
67 "status": "Generated"
68 },
69 {
70 "filename": "transformer_family.png",
71 "purpose": "Slide 16 hero — faded Transformer descendants tree backdrop (root + two intermediate branches + many small leaf shapes); calm middle band reserved for SVG closing title",
72 "page_role": "hero_page",
73 "text_policy": "none",
74 "aspect_ratio": "16:9",
75 "image_size": "2K",
76 "prompt": "Technical blueprint schematic style. Clean precise lines on a very subtle implied grid. Elements are simplified to schematic forms — rounded rectangles and small abstract leaf shapes connected by thin straight lines. Crisp uniform stroke around 1.5px. No textures, no shading, no painterly artifacts; engineering-precise. The entire schematic is rendered at low opacity around 35% so it reads as a faded backdrop. Color behavior is restrained-corporate: secondary off-white #F5F7FA dominates the canvas as breathing space (about 68% area); primary deep navy #1A365D draws the root rectangle, the two intermediate rectangles, the descendant leaf shapes, and all connector lines (about 27% area, all rendered at ~35% opacity to fade); accent blue #3182CE outlines only the root rectangle and one selected leaf in full opacity (under 5% area), marking the genealogical anchor and one focal descendant; no fourth color. A subtle 6% opacity navy grid sits beneath everything. Custom composition — a symmetric top-down tree-of-descendants schematic occupies the upper third and lower third of the canvas with the middle horizontal band of the canvas (roughly 30% of canvas height across the full width) deliberately calm as breathing space reserved for a closing title overlay in SVG. Tree structure: at the top center, one single rounded rectangle slightly larger than the rest (the root Transformer) outlined in accent blue at full opacity. From the root, thin lines fan downward across the calm middle band and emerge in the lower third, branching to two intermediate rounded rectangles set symmetrically left and right (suggesting encoder-only and decoder-only families). From each intermediate, three to four thinner branches drop further down to small abstract leaf shapes of varying small sizes (suggesting the multitude of derived models). All elements other than the root and one selected leaf are rendered in primary navy at 35% opacity. Reads as a quiet 'what this paper started' diagram. Composed as a 1280×720 PPT 16:9 hero background; the middle horizontal band must remain calm for SVG closing title overlay. NO text, letters, numbers, model names, signs, watermarks, or written symbols anywhere in the image. Color values (HEX codes like #1A365D) and color names are rendering guidance only — do NOT display HEX codes, color names, or palette labels as visible text anywhere in the image.",
77 "alt_text": "Faded blueprint tree of Transformer descendants with accent-blue root and one accent leaf, calm middle band reserved",
78 "status": "Generated"
79 }
80 ]
81 }
82
82 lines JSON