| name | research-ppt-imagegen |
| description | Use when the user wants PPT slide images or imagegen-generated presentation pages from a thesis, graduation thesis, paper, article, literature review, tables, research notes, reference PPT/images, or asks for 答辩PPT图片版, 每页PPT讲什么/画什么, PPT内容框架图, 论文生成PPT, image-2, gpt-image-2, or simplaj imagegen slide visuals. |
Research PPT Imagegen
Core Principle
Turn research materials into polished 16:9 academic slide images by building the page logic first, then generating each slide image from a concrete page blueprint. The target is a dense, defense-ready research presentation page, not a poster, stock-image collage, or copied paper figure.
Use With Other Skills
- Use
image2-router for all image-2 / gpt-image-2 / simplaj imagegen generation or editing.
- If the user provides
.pdf, .pptx, .docx, .xlsx, .csv, images, or a project folder, use the matching file skill or local extraction tools before planning slides.
- If the user provides a reference slide/image, inspect it for layout, color, typography, and density; adapt the style without blindly copying irrelevant content.
Input Reading
First inventory what the user actually provided:
- thesis or paper: title, problem, method, contributions, figures, tables, experiments, numbers, conclusion;
- literature/article: research question, main method, architecture, results, limitations, discussion points;
- tables/results: metrics, comparisons, best rows, deltas, conditions, caveats;
- reference PPT/images: page ratio, title bar, grid, panel style, colors, font feel, chart/diagram style;
- existing
pre.md, defense_ppt_plan.md, or slide outline: treat it as the starting plan and improve it only where needed.
Ask only for critical missing information. If the user wants an answer now, choose sensible defaults: Chinese academic defense, 10 minutes, 10 main slides plus optional backups, mixed expert/non-expert audience, professional blue-white technology style.
Mode Selection
Before planning, classify the task into one primary mode. If multiple modes fit, choose the one that matches the user's deliverable, not just the source file type.
thesis-defense: graduation thesis, defense PPT, 10-minute oral defense, committee presentation.paper-reading: single paper/literature reading, journal club, group meeting paper report.literature-review: multiple papers, related work, survey, method taxonomy, comparison across works.experiment-report: tables, metrics, ablations, benchmark results, data-analysis slides.project-progress: project update, midterm report, current progress, issues, next steps.single-slide: one specific page image or revision, usually from an existing plan or reference.full-deck-images: generate images for every planned slide.
Default mode rules:
- Mentions of
毕业论文, 答辩, 本科/硕士/博士论文, or 10分钟 -> thesis-defense.
- A single article/paper for discussion ->
paper-reading.
- Several papers or "综述/相关工作/文献调研" ->
literature-review.
- Mostly tables/results/metrics ->
experiment-report.
- Existing
第 N 页, slide N, or a provided reference slide -> single-slide, but create/read the deck plan first if the page is undefined.
Task Contract
At the start of work, establish a compact task contract. This can be internal for obvious requests, but make it explicit when the user may be mixing planning, slide selection, provider choice, and image generation.
Mode:
Input materials:
Delivery:
Slide scope:
Audience/time:
Style source:
Provider:
Output paths:
Next action:
Allowed Delivery values:
plan-only: produce the overall PPT framework and per-slide construction plan; do not generate images.slide-plan-only: produce or revise the construction plan for selected slides only.single-slide-image: generate one slide image from an existing or newly inferred plan.plan-plus-single-image: create/read the full plan, then generate one selected slide.full-deck-images: create/read the plan, then generate every slide image.revision-image: revise an existing generated slide image using the user's critique.
Task contract rules:
- Provider names matter only when
Delivery includes image generation. They are not needed for plan-only or slide-plan-only.
- If the user names a provider, keep it:
simplaj provider -> --provider simplaj; fhl imagegen -> --provider fhl.
- If no provider is named and images are needed, use
image2-router automatic routing.
- If the user says
不要生图, 先不要生成图片, or asks only for a framework, set Delivery: plan-only.
- If the user asks for
第 N 页 but no page plan exists, first create or infer the overall slide plan so that page has a concrete title, purpose, source evidence, and layout. Do not blindly invent an isolated slide.
- If the user provides an existing
pre.md, defense_ppt_plan.md, slide_plan.md, or reference PPT, use it as the plan/style source before inventing a new structure.
Typical commands:
- Planning only:
用 research-ppt-imagegen 读取这篇论文，先生成 10 分钟答辩 PPT 的整体框架和每页施工图，不要生图。
- Single slide from an existing plan:
用 research-ppt-imagegen，用 simplaj provider，根据 defense_ppt_plan.md 生成第 3 页图片版。
- One-shot plan plus image:
用 research-ppt-imagegen，用 simplaj provider，先根据我的论文生成整体框架和每页施工图，然后只生成第 3 页图片版。
- Full deck images:
用 research-ppt-imagegen，用 simplaj provider，根据这个框架生成所有页面的图片版。
Deck Brief And Narrative
For plan or deck-level work, write a brief before drafting slides:
Audience:
Goal:
Duration:
Language:
Deck Type:
Expected Outcome:
Required Sections:
Visual Risk Points:
Choose one narrative shape that matches the selected mode:
thesis-defense: problem -> gap -> approach -> evidence -> conclusion.paper-reading: problem -> method -> experiment -> limitation -> takeaway.literature-review: field map -> method families -> representative works -> comparison -> trend.experiment-report: setup -> metric -> main result -> analysis -> conclusion.project-progress: goal -> progress -> blockers -> decisions -> next steps.
If time is short, merge detail slides before deleting narrative bridge slides. Lost transitions make a technical talk harder to follow than one fewer detail page.
Page Plan Before Imagegen
Do not start image generation before a per-slide "施工图" exists. For each slide, write or update a compact plan with:
Slide XX
Title:
One-sentence message:
Source evidence:
Source location:
Narrative role:
Transition:
Visible text:
Visual composition:
Reusable assets:
New elements to draw:
Speaker point:
Time:
Image prompt notes:
The plan must say what the slide should prove, what evidence it uses, and what visual structure should be drawn. Avoid vague entries like "show method" or "show results"; specify flowcharts, cards, tables, callouts, result strips, timelines, or comparison matrices.
Figure Intent Cards
Before generating a complex slide visual, write a compact intent card for each major diagram, chart, table, or evidence block:
Purpose:
Figure type: diagram | data-chart | comparison | timeline | table | callout | result-strip
Required components or data:
Exact labels:
Preferred layout:
Must-preserve facts:
Style cues:
Output path:
Use the cards to prevent image generation from inventing modules, arrows, labels, metrics, or conclusions. If a slide contains several visual units, keep each card short and use the slide composition to explain how they fit together.
Default Defense Structure
For a 10-minute thesis defense, prefer this shape unless the user gives a better one:
- Cover: thesis title, author, advisor, institution, restrained academic visual.
- Background and challenge: why the problem matters; one problem chain or gap diagram.
- Research question and contribution: 2-3 contributions with evidence anchors.
- Overall framework: end-to-end pipeline or system/method overview.
- Modeling or representation: key abstraction, variables, constraints, or semantic mapping.
- Method details: search/generation/optimization/data-processing flow with decision points.
- Evaluation setup: datasets, baselines, metrics, experiment matrix.
- Main results: numeric comparison, best result highlight, concise takeaway.
- Analysis or ablation: why the result happens; bottleneck, tradeoff, or sensitivity view.
- Summary and outlook: conclusion, limitations, future work.
Optional section directory pages can use: 01 背景与问题 / 02 建模与表达 / 03 评估与结果 / 04 总结与展望. Keep directory slides light but still visual.
Slide Image Style
Use these defaults unless the user supplies another visual direction:
- 16:9 full-slide PNG, usually
2048x1152 or another 16:9 size accepted by the provider.
- Academic defense / research presentation, clean Chinese sans-serif feel like Microsoft YaHei or Noto Sans CJK.
- White or very light background, restrained blue/teal/gray accents, one secondary warm accent only for highlights.
- Clear title bar, section marker, dense but readable layout, aligned grid, consistent margins.
- Prefer diagrams, cards, matrices, small tables, result strips, axis-like charts, and callouts over decorative art.
- Use thesis/paper figures as local evidence or cropped support, not as the whole slide.
- Include exact numbers, metric names, method names, and labels from the source when they matter.
Hard Visual Rules
- No stock-photo collage, overlapping screenshots, random image stacking, or "图片叠加感".
- No slide that is only a pasted thesis figure or paper screenshot.
- No icon spam. Use icons only when they carry meaning; small decorative icons should be removed.
- No marketing hero page, gradient-only poster, empty "AI presentation" template, or generic technology background.
- No unreadable fake paragraphs, garbled Chinese, watermark, QR code, fake logos, or tiny text walls.
- No one-note palette dominated by a single saturated hue.
- No clutter that makes the main message ambiguous.
Chinese Text Policy
Image models can garble long Chinese text. Keep visible generated text short and intentional:
- title: 8-18 Chinese characters when possible;
- panel labels: 2-8 Chinese characters;
- callouts: one short phrase each;
- tables: few columns, short headers, exact numeric values;
- avoid full paragraphs inside the image.
If the slide needs a lot of exact Chinese text, generate a clean diagram/background with short labels and plan to overlay final text in PPT/native tools later. For image-only delivery, still request crisp Simplified Chinese and inspect the output carefully.
Chart And Diagram Routing
Choose the visual production path before writing the image prompt:
- Numerical axes, exact values, ablations, rankings, heatmaps, or reproducible comparisons -> prefer deterministic chart generation with matplotlib/seaborn or a native chart asset, then use imagegen only for surrounding slide composition if needed.
- Conceptual pipeline, system architecture, taxonomy, problem chain, or page-level visual layout -> use image-2 through
image2-router.
- Exact but compact result communication -> use numeric cards, small tables, and result strips instead of asking image-2 to invent full chart axes.
- Reference figure redraw -> use image-to-image/edit only after reading the surrounding source context; captions alone are not enough when factual structure matters.
When a chart or label must be exact, do not rely on unchecked bitmap text. Generate the chart as a separate asset or plan native PPT text overlay.
Prompt Template
Write one prompt file per slide under tmp/imagegen/ or the project-local equivalent:
1. FRAMING
Create one polished 16:9 academic research presentation slide image.
Canvas: 2048x1152 px, full-slide PNG, no border, no watermark.
Deck mode: <mode>
Audience and purpose: <audience / goal>
2. VISUAL STYLE
Style: Chinese thesis defense or selected style preset, clean blue-white technology style, high information density, restrained colors, Microsoft YaHei / Noto Sans CJK feel, crisp alignment.
3. COLOR PALETTE
Use a consistent deck palette: white/light background, deep blue/teal/slate accents, one warm highlight color only for key results.
4. LAYOUT
Slide section: <section name>
Slide title: <exact title>
Core message: <one sentence>
Required visible text:
- <short title / labels / numbers only>
Source facts to reflect:
- <facts, metrics, figure names, table values, method names>
Composition:
- <precise layout: e.g. top title bar; left problem chain; center pipeline; right result cards; bottom takeaway strip>
- <which source figure/table is represented and how: cropped evidence, simplified chart, callout, etc.>
- <new visual elements to draw: flowchart, cards, arrows, matrix, result strip>
5. CONNECTIONS / DATA MAPPING
- <every major arrow, source -> target, label, and flow direction>
- <how listed numbers/tables map to visible cards, tables, chart assets, or callouts>
6. CONSTRAINTS
- dense but readable academic slide;
- all panels aligned to a clean grid;
- use exact numeric values where listed;
- avoid decorative icon spam;
- avoid stock-photo collage or stacked-image look;
- no garbled Chinese text, no fake long paragraphs, no watermark.
For a revision prompt, name the defect directly:
Regenerate the same slide with higher information density. Keep the title and core layout, but add a compact result table and two explanatory callout cards. Remove most small icons. Make the page look like a coherent designed slide, not layered images.
Image2 Generation
Use image2-router commands. If the user names a provider, keep it; for example, "simplaj imagegen" means --provider simplaj. If not named, let the router auto-select.
python "$HOME/.codex/skills/image2-router/scripts/image2_router.py" generate \
--provider simplaj \
--prompt-file tmp/imagegen/slide03_prompt.txt \
--size 2048x1152 \
--out output/imagegen/slide03.png
For reference-based style or iterative correction, use edit with the old slide or reference image:
python "$HOME/.codex/skills/image2-router/scripts/image2_router.py" edit \
--provider simplaj \
--image output/imagegen/slide03.png \
--prompt-file tmp/imagegen/slide03_revision_prompt.txt \
--out output/imagegen/slide03_v2.png
Store outputs in the project when possible:
<project>/ppt_assets/slideXX_<short_name>.png
<project>/tmp/imagegen/slideXX_prompt.txt
<project>/defense_ppt_plan.md or <project>/slide_plan.md
Inspection Loop
After each generated slide, inspect the image before moving on. Review in three layers:
- Narrative check: the slide has one clear job, fits the selected narrative shape, and transitions from or to adjacent slides.
- Visual check: the slide looks like a coherent academic PPT page, not a poster, collage, or copied figure.
- Fact check: labels, arrows, numbers, metrics, and conclusions match the source evidence or page plan.
Regenerate or edit if any of these appear:
- information density is too low for an academic defense;
- slide is mostly decoration or a single large paper figure;
- too many small icons;
- Chinese text is garbled or visually fake;
- layout looks like pasted screenshots/photos;
- source facts or numbers are wrong;
- title, section, or main message is missing;
- chart/table labels are too small to read.
When the user says "信息密度再高一些", add meaningful content: numeric cards, compact comparison tables, flow steps, conditions, annotations, and bottom takeaway strips. Do not solve it by adding random icons or tiny filler text.
When the user says "图片叠加感太重", rebuild the prompt around one coherent layout grid. Replace photo/screenshot layers with native-looking cards, simplified diagrams, clean chart blocks, and consistent panel backgrounds.
Delivery
Return the saved image paths and a short note on what was generated or revised. If many slides are produced, include a compact slide list with filenames. Mention any limitations honestly, especially if a generated image has text fidelity risk and should later receive native PPT text overlay.
Common Mistakes
- Summarizing the paper without designing the slide.
- Generating images before extracting evidence and page intent.
- Treating a thesis figure as the whole page.
- Asking image-2 for long Chinese paragraphs.
- Adding many icons to fake density.
- Making a pretty poster that lacks research content.
- Changing providers when the user explicitly named one.
- Delivering without visually checking the generated PNG.
