PaperBanana - Academic Illustration Generator

Multi-agent pipeline (Retriever → Planner → Stylist → Visualizer → Critic) for publication-quality academic diagrams, statistical plots, and presentation slides.

API key: Set provider keys in PaperBanana project's .env file. Timeout: 300000 (5 min) for all generation commands.

---

Commands

All commands run from project root: cd <paperbanana_dir> && python -m paperbanana.cli <cmd>

Command Selection Decision Tree

Route user requests to the right subcommand before looking up parameters:

User intent	Signal words	Subcommand
方法论/架构/流程图 from text or PDF	"method figure", "架构图", "流程图", "methodology", "pipeline diagram", "论文配图"	generate
Statistical plot from data file	"plot", "curve", "bar chart", "scatter", "heatmap", has CSV/JSON	plot
Single presentation slide	"slide", "一张幻灯片", "封面图", single prompt file	slide
Batch slide generation	"all slides", "批量生成", "N 张幻灯片", prompts/ directory	slide-batch
Compare generated vs human reference	"evaluate", "对比", "与参考图对比"	evaluate
Manage reference dataset	"download dataset", "清缓存"	data
First-time provider config	"setup", "配置 API key"	setup

Ambiguous input: If user provides just a description with no subcommand signal, default to generate (see Argument Parsing table for details).

Out-of-scope: Pure code generation (matplotlib/seaborn script) is NOT paperbanana's job — those go to matplotlib / scientific-visualization skills. Paperbanana is for AI-driven image generation + critique loops.

Note (upstream sync pending): Upstream paperbanana CLI also adds subcommands (plot-batch #123, sweep #118) not yet reflected in this table. See the llmsresearch/paperbanana CHANGELOG for the authoritative CLI surface.

generate — Methodology Diagrams

python -m paperbanana.cli generate --input '<file>' --caption '<caption>' --optimize --verbose

When user provides inline text (no file): write to temp file, use as --input.

Parameter	Default	Description
--input / -i	—	Path to methodology text file or PDF (.pdf requires pip install 'paperbanana')
--caption / -c	—	Figure caption / communicative intent
--output / -o	auto	Output image path
--vlm-provider	gemini	VLM provider: gemini, anthropic, openai, bedrock, openrouter, ollama, claude_code, litellm
--vlm-model	auto	VLM model name
--image-provider	auto	Image gen provider: google_imagen, openai, bedrock, openrouter
--image-model	auto	Image gen model name
--iterations / -n	3	Max critic rounds
--auto	off	Loop until critic is satisfied (safety cap via --max-iterations)
--max-iterations	30	Safety cap for --auto mode
--optimize	off	Preprocess inputs (parallel enrichment + caption sharpening)
--continue	off	Continue from the latest run
--continue-run	—	Continue from a specific run ID
--feedback	—	User feedback for the critic when continuing a run
--aspect-ratio / -ar	auto	Target aspect ratio: 1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, 21:9
--format / -f	png	Output format: png, jpeg, webp
--dry-run	off	Validate inputs without making API calls
--exemplar-retrieval	off	Enable external exemplar retrieval before planning
--seed	—	Random seed for reproducible generation
--verbose / -v	off	Show detailed agent progress and timing
--auto-download-data	off	Auto-download expanded reference set (~257MB) on first run
--venue	—	Academic venue style: neurips, icml, acl, ieee, custom
--pages	—	Page range for PDF input (e.g., 3-5)
--config	—	Path to config YAML file

Venue styles: --venue neurips applies NeurIPS-specific methodology and plot style guides from data/guidelines/. Each venue has distinct color palettes, layout conventions, and typography expectations.

PDF input: --input paper.pdf --pages 3-5 extracts text from the specified pages as source context.

Exemplar advanced flags: --exemplar-retrieval enables retrieval; see generate --help for additional config flags (--exemplar-endpoint, --exemplar-mode, --exemplar-top-k, --exemplar-timeout, --exemplar-retries).

plot — Statistical Plots

python -m paperbanana.cli plot --data '<data.csv>' --intent '<intent>' --optimize --verbose

Parameter	Default	Description
--data / -d	—	Path to data file (CSV or JSON) [required]
--intent	—	Communicative intent for the plot [required]
--output / -o	auto	Output image path
--vlm-provider	gemini	VLM provider
--iterations / -n	3	Refinement iterations
--format / -f	png	Output format
--aspect-ratio / -ar	auto	Target aspect ratio
--optimize	off	Enrich context and sharpen caption
--auto	off	Loop until critic satisfied
--verbose / -v	off	Detailed progress

slide — Presentation Slides

python -m paperbanana.cli slide --input '<prompt.md>' --resolution 4k

Parameter	Default	Description
--input / -i	—	Path to slide prompt markdown file [required]
--caption / -c	auto	Slide intent description
--output / -o	auto	Output image path
--image-model	auto	Image gen model
--vlm-model	auto	VLM model name
--iterations / -n	3	Max critic rounds
--style / -s	—	Style preset name (see table below)
--list-styles	off	List all available style presets and exit
--resolution / -r	4k	Output resolution: 1k, 2k, 4k
--config	—	Path to config YAML file

slide-batch — Batch Slide Generation

python -m paperbanana.cli slide-batch --prompts-dir '<dir>' --resolution 4k

Parameter	Default	Description
--prompts-dir	—	Directory containing slide prompt markdown files [required]
--output-dir	auto	Output directory
--image-model	auto	Image gen model
--style / -s	—	Style preset applied to all slides
--iterations / -n	3	Max critic rounds per slide
--resolution / -r	4k	Output resolution
--concurrent / -c	2 (settings.batch_concurrent)	Slides generated concurrently; 3 is the sweet spot, never exceed 4. Requires a paperbanana build ≥ 2026-08-03 (maintainer's fork)

Wave-Parallel Batch Generation (speed default for ≥2 slides)

With a concurrency-enabled paperbanana build, batch generation runs slides in parallel with identical per-slide quality — every slide keeps its full Critic loop, its own pipeline instance, and its own run directory:

python -m paperbanana.cli slide-batch --prompts-dir '<dir>' --output-dir '<out>' --resolution 4k --concurrent 3

Measured (2026-08-03): 6 slides at --concurrent 3 in 309s vs 768s serial estimate (0.40x, ~2.5x speedup). Built-in resilience: 5s start-up stagger (same-second bursts to the image API fail or hang server-side long before per-minute quotas are near), in-batch delayed retry for transient 503s (recovery overlaps with other slides), and an end-of-batch serial retry pass for stragglers. Delivery quality is protected twice over: the final image per slide is the highest-critic-score iteration (not simply the last), and critic_score_threshold=9.0 skips provably-done rounds early — calibrated on 69 historical runs with zero false early-stops.

If the installed paperbanana lacks --concurrent, fall back to serial slide-batch — do NOT spawn more than 3 parallel slide processes yourself, as there is no cross-process rate-limit coordination.

Style Presets (23 available)

Use --style <name> with slide or slide-batch. Use --list-styles to see all.

Style	Source	Best For
blueprint	baoyu	Architecture, system design, technical
chalkboard	baoyu	Classroom, teaching, education
corporate	baoyu	Business, investor, quarterly reports
minimal	baoyu	Executive briefings, clean/simple
sketch-notes	baoyu	Tutorials, guides, beginner content
watercolor	baoyu	Lifestyle, wellness, artistic
dark-atmospheric	baoyu	Entertainment, gaming, cinematic
notion	baoyu	SaaS, product, dashboards
bold-editorial	baoyu	Product launches, keynotes, marketing
editorial-infographic	baoyu	Science communication, explainers
fantasy-animation	baoyu	Storytelling, magical, children
intuition-machine	baoyu	Academic research, bilingual
pixel-art	baoyu	Gaming, retro, developer culture
scientific	baoyu	Biology, chemistry, medical
vector-illustration	baoyu	Creative, children, flat design
vintage	baoyu	Historical, heritage, expedition
tech-keynote	elite-ppt	Apple/Tesla premium minimalism
creative-bold	elite-ppt	Google/Airbnb energetic innovation
financial-elite	elite-ppt	Goldman Sachs/McKinsey sophistication
biotech	sci-slides	Life sciences, genomics
neuroscience	sci-slides	Brain research, cognitive science
ml-ai	sci-slides	Machine learning, deep learning
environmental	sci-slides	Ecology, climate, sustainability

evaluate — Comparative Evaluation

python -m paperbanana.cli evaluate --generated '<gen.png>' --reference '<ref.png>' \
  --context '<text_file>' --caption '<caption>'

Parameter	Default	Description
--generated / -g	—	Path to generated image [required]
--reference / -r	—	Path to human reference image [required]
--context	—	Path to source context text file [required]
--caption / -c	—	Figure caption [required]
--vlm-provider	gemini	VLM provider for evaluation
--verbose / -v	off	Detailed progress

data — Manage Reference Datasets

python -m paperbanana.cli data download   # Download expanded reference set (~257MB)
python -m paperbanana.cli data info       # Show cached dataset info
python -m paperbanana.cli data clear      # Remove cached dataset

ablate-retrieval — Retrieval Ablation (Advanced)

Research utility for running baseline vs retrieval ablation (k sweep). See ablate-retrieval --help for details.

setup — Interactive Setup Wizard

python -m paperbanana.cli setup

Guides through API key configuration and provider selection. No flags needed.

---

Provider Selection

Provider	VLM	Image Gen	Setup
Google Gemini	Flash / Pro	Imagen 3	GOOGLE_API_KEY
Anthropic Claude	Claude 4	—	ANTHROPIC_API_KEY
OpenAI	GPT-4o	DALL-E 3	OPENAI_API_KEY
AWS Bedrock	Claude / Nova	Nova Canvas	AWS credentials
OpenRouter	Various	Various	OPENROUTER_API_KEY
LiteLLM	100+ backends	via backend	LITELLM_MODEL / LITELLM_API_KEY
Ollama	Local models	—	OLLAMA_BASE_URL / OLLAMA_MODEL
Claude Code	via claude CLI	—	Claude Code signed in (no key)

Use --vlm-provider and --image-provider flags to select providers per command.

⚠️ Provider naming asymmetry (common trap): VLM providers use short names (gemini / openai / anthropic / bedrock / openrouter), but image providers require the _imagen suffix: google_imagen (not gemini), openai_imagen (not openai), bedrock_imagen, openrouter_imagen. The error ValueError: Unknown image provider: openai. Available: google_imagen, openrouter_imagen, openai_imagen, bedrock_imagen means you hit this. Every --image-provider example in this document uses the X_imagen form.

Auto-routing decision table

When the user doesn't specify an image model, pick --image-provider by the first matching rule (top priority wins). Calibrated from a controlled 16-prompt two-provider comparison (2026-04) plus slide-deck production use:

#	Condition	Flag to pass	Why
1	User explicitly names a provider ("use GPT" / "use Gemini" / "nano banana")	That provider	Explicit override beats all auto-rules
1	User asks for a side-by-side ("both providers", "对比一下")	Run once with each, show both	Dual comparison
2	Submission-ready venue figures ("投稿用", journal/conference figure)	--image-provider openai_imagen --image-model gpt-image-2	Rigor and text fidelity priority
3	slide / slide-batch AND the prompt contains CJK text	--image-provider openai_imagen --image-model gpt-image-2	Avoids a documented Gemini duplicate-character bug on CJK slide titles
3	≥2 reference images for editing/composition	--image-provider openai_imagen --image-model gpt-image-2	Avoids Gemini multi-image hallucination (invented text fields)
4	Traditional East-Asian aesthetics (水墨 / 书法 / 古风 / ukiyo-e etc.)	--image-provider google_imagen	Gemini dominates this space (3-0 sweep in comparison test)
4	generate with dense multi-module figures (architecture / ablation / encoder-decoder)	--image-provider openai_imagen --image-model gpt-image-2	GPT wins on structural preservation
5	Everything else	--image-provider google_imagen (default)	Better aesthetics on ordinary tasks, ~2x faster, cheaper

Never second-guess an explicit user choice. If the user says which provider to use, use it — even when the auto-rules would pick differently.

Argument Parsing

Input Pattern	Action
generate <file.txt> <caption>	Generate with file + caption
generate <text>	Write to temp file, auto-caption, generate
plot <data.csv> <intent>	Plot mode
slide <prompt.md>	Generate presentation slide
slide-batch <dir>	Batch generate slides
evaluate <gen.png> <ref.png>	Comparative evaluation
Just a description (no subcommand)	Default to generate

Error Handling

Two types of API failure can occur during generation. Handle them differently:

Type 1: Image Generation API Failure (Visualizer)

The image provider (Gemini Imagen, DALL-E, Nova Canvas) fails to return an image.

Error	Cause	Action
429 / ResourceExhausted	Rate limit	Wait 30s, retry up to 3 times
500 / 503 / ServerError	Provider outage	Switch to fallback provider (see chain below)
400 / InvalidArgument	Bad prompt (too long, policy violation)	Shorten/rephrase prompt, retry once
401 / 403	Invalid API key	Stop and ask user to run setup
Timeout (>60s no response)	Network or provider hang	Retry once, then switch provider

Fallback chain: google_imagen → openai → bedrock → openrouter. Use the next provider in chain that has a valid API key in .env. If all fail, stop and report the error.

Type 2: VLM Critic API Failure

The VLM provider (Gemini Flash, Claude, GPT-4o) fails during quality evaluation.

Error	Cause	Action
429 / Rate limit	Too many requests	Wait 15s, retry up to 3 times
JSON parse failure	VLM returned malformed response	Do NOT treat as "approved". Retry once with stricter prompt. If still fails, mark output as UNREVIEWED
500 / 503	Provider outage	Switch VLM provider (see chain below)
Timeout (>30s)	Network hang	Retry once, then skip Critic and mark as UNREVIEWED

VLM fallback chain: gemini → anthropic → openai → openrouter.

Critical rule: A Critic failure must NEVER silently approve an image. If Critic cannot evaluate, the output status must be UNREVIEWED, not APPROVED. Report this clearly to the user.

Recovery with --continue

Use --continue to resume after any failure:

Scenario	Command
Pipeline crashed mid-generation	--continue (resumes latest run)
Want to iterate on a specific run	--continue-run <run_id>
Want to provide feedback for next iteration	--continue --feedback "make the arrows thicker"

The run directory preserves all intermediate state (plans, images, critic feedback). --continue picks up from the last successful step.

Batch Mode (slide-batch) Resilience

When generating multiple slides, a single slide failure should NOT kill the batch:

1. Log the failure for the specific slide
2. Continue generating remaining slides
3. At the end, report which slides succeeded and which failed
4. User can re-run with --continue to retry only failed slides

---

🔴 User Confirmation Checkpoints

Paperbanana is CLI-first, but three user-facing actions are expensive or irreversible. 🛑 STOP and ask for explicit confirmation before running any row below — do not proceed on assumed consent.

🔴 Trigger	🛑 STOP — Confirm before proceeding
--auto with --max-iterations > 5	🛑 STOP. Show: cap, est. API cost (≈ iterations × $0.04), est. wall time (≈ iterations × 30s). Ask: "Proceed with up to N iterations?" Do not kick off until user says yes.
--auto-download-data on first run	🛑 STOP. Announce: "reference dataset will be downloaded to cache (~257MB full_bench, or lightweight curated set in upstream ≥ #112)". Ask: "Continue?" Do not download until confirmed.
setup wizard	🛑 STOP. Before writing to .env, show the exact keys and preview of values (redact secrets after 4 chars). Ask: "Save to .env?" Do not write until confirmed.

✅ No checkpoint needed: normal generate / plot / slide (no --auto, within iteration cap 3) — these are short, cheap, and the Critic loop is self-bounded. Run them directly.

---

⛔ Anti-Patterns — Red-Line Blacklist

Hard "do NOT" rules. Each maps to a failure mode already encoded above — this section consolidates them into one scannable list.

⛔ Anti-pattern	Why it's wrong	Do this instead
Treating a Critic API failure as "approved"	Ships an unreviewed image as if it passed QA	Mark UNREVIEWED, never APPROVED; report to user
Proceeding past a 🔴 checkpoint without confirmation	Burns API budget / overwrites .env on assumed consent	🛑 STOP at every checkpoint row; wait for explicit yes
Writing matplotlib/seaborn scripts	That's scientific-visualization's job, not paperbanana	Route code-gen away; paperbanana = AI image gen + critique loop
Killing the whole slide-batch on one slide failure	Loses N−1 good slides over 1 bad one	Log the failure, continue, report survivors, retry via --continue
Full regeneration after a mid-run crash	Throws away plans/images/critic state, wastes API spend	Resume with --continue / --continue-run <id>
Routing every "make a figure" to generate	plot (data files) and slide (presentation) have dedicated paths	Run the Command Selection Decision Tree first
Inventing CLI flags not in the parameter tables	Upstream CLI surface drifts (see #115/#118/#123 note)	Verify with <cmd> --help; don't fabricate flags

---

After Generation

1. Parse output to find image path
2. Use Read tool to display the generated image
3. Report Run ID, iteration count, and Critic feedback
4. If any outputs are marked UNREVIEWED, warn the user explicitly
5. If user expresses dissatisfaction OR status is UNREVIEWED, proactively suggest: python -m paperbanana.cli <cmd> --continue --feedback "<specific fix>" — preserves run state, avoids full regeneration