paper-to-image

---

name: paper-to-image description: Turn any research paper (PDF, arXiv URL, .docx, HTML, .tex export) into publication-ready figures via a 5-step pipeline — read the paper, write an in-depth résumé with a figure catalog, choose a render path (code for charts/diagrams/chat UIs; Nano Banana JSON for conceptual art), then generate PNGs. Handles three figure archetypes found in systems/ML papers: chat-UI mockups, architecture diagrams, and performance charts. Default output looks like the actual figures inside a real paper. Use when the user asks for paper figures, diagrams, schematics, system overviews, result charts, or illustrated conversation examples from a paper, abstract, or scientific PDF.

Paper → Summary → Figure Plan → Code or Prompt → PNG

A generic, paper-agnostic pipeline. Five sequential steps. Do not skip any. Each step's output feeds the next.

Paper ──▶ Markdown text ──▶ In-depth résumé + figure catalog
       ──▶ Render-path routing (code vs Nano Banana)
       ──▶ Python script or JSON prompt
       ──▶ PNG in output/generated/

The skill ships with a worked reference implementation (the MemGPT paper, all 8 figures) under scripts/render_all_memgpt.py. Treat it as a calibration target — once you can reproduce its figures, the same primitives produce comparable figures for any other systems/ML paper.

Project layout

input/                          ← drop the source paper(s) here
output/
  benchmark/                    ← (optional) reference figures to match exactly
  generated/                    ← timestamped renders end up here
scripts/
  render_helpers.py             ← shared palette + drawing primitives (import these)
  render_paper_template.py      ← starter scaffold for a NEW paper
  render_all_memgpt.py          ← worked reference example
  setup.py / generate.py        ← provider config + Nano Banana dispatcher
SKILL.md                        ← this file

First-run setup

Check if the provider is configured:

test -f config.json && python3 -c "import json; c=json.load(open('config.json')); print('Provider:', c['provider'], '|', 'Model:', c.get('model','n/a'))" || echo "Not configured"

If not configured — or if the user wants to change provider — run the interactive setup:

python3 scripts/setup.py

This asks:

1. Which provider: Gemini (default), fal.ai, or Replit platform.
2. Which model within that provider.
3. Checks whether the required API key env var is set.

Writes config.json (gitignored). All subsequent renders read from it automatically.

See references/providers.md for API key setup, model lists, and per-provider caveats.

✋ Stop here if the provider is not yet configured and the paper has any nano-banana figures. Code-rendered figures don't need a provider.

Workflow checklist

Copy this checklist into the chat and tick items as you progress:

- [ ] Setup: provider configured (only needed for nano-banana figures)
- [ ] Step 1: Read the paper (convert to Markdown)
- [ ] Step 2: Write in-depth résumé + figure catalog
- [ ] Step 3: Choose render path for each figure
- [ ] Step 4a (code path): Write & run Python script
- [ ] Step 4b (Nano Banana path): Build JSON prompt and validate
- [ ] Step 5: Generate / save to output/generated/
- [ ] (Optional) Compare to output/benchmark/ and iterate

---

Step 1 — Read the paper

Convert any source to Markdown so the full text is available. Use uvx markitdown or pdftotext.

uvx markitdown input/<paper>.pdf -o input/<paper>.md
# or
pdftotext input/<paper>.pdf input/<paper>.txt

Notes:

• Works for PDF, .docx, .pptx, HTML, EPub, and more.
• For arXiv links, prefer the PDF URL (https://arxiv.org/pdf/<id>.pdf).
• For scanned PDFs, re-run with Azure Document Intelligence (-d -e ENDPOINT).
• Read the resulting .md / .txt before continuing.

✋ Stop here if the paper text is too short (< 500 words) or you cannot extract clean text.

---

Step 2 — In-depth résumé + figure catalog

Write the résumé directly in chat (or save to summary.md). Use this outline:

# <Paper title> — In-depth résumé

## Problem & motivation
What gap or question the paper addresses and why it matters.

## Method
Core technique, architecture, formalism, or experimental design. Include
notable equations, diagrams, or pseudocode the paper relies on.

## Key results
Quantitative findings, tables, headline numbers, ablations.
**Extract exact numbers from tables and result sections** — these will feed
chart figures directly.

## Limitations & open questions
Failure modes, caveats, and what the authors flag as future work.

## Visual hooks
3–5 concrete image-able concepts from the paper.

## Figure catalog
For every figure in the paper, list one entry:

| # | Description | Type | Render path | Notes |
|---|-------------|------|-------------|-------|
| 1 | <what the figure shows> | <type — see table below> | code \| nano-banana | dimensions, key colors, key labels |

Figure types and their render paths:

Type	When to use	Render path
chat_mockup	Annotated chat/conversation UI screenshots	code
architecture_diagram	System block diagrams with colored components	code
line_chart	Multi-line accuracy/loss vs X-axis plots	code
bar_chart	Category comparisons on one metric	code
scatter_plot	Point clouds with labeled axes	code
schematic_flow	Pipeline box-and-arrow (few nodes, B&W LaTeX)	nano-banana
conceptual	Metaphor, cover art, concept illustration	nano-banana
venn / quadrant	Positioning vs prior work	nano-banana

✋ Stop here if you cannot find at least one concrete visual hook. A weak hook produces a generic image; ask the user which aspect of the paper they want illustrated.

---

Step 3 — Choose render path

For each figure in the catalog, pick a path:

• Code path (chat_mockup, architecture_diagram, line_chart, bar_chart, scatter_plot) — write a Python script using matplotlib and run it directly. Always faster, always typographically exact.
• Nano Banana path (schematic_flow, conceptual, venn, quadrant) — build the JSON prompt and use scripts/generate.py.

When in doubt, prefer the code path. Diffusion models (including Gemini 3 Pro Image) frequently mangle long labels, made-up acronyms, and equations. Code is exact.

---

Step 4a — Code-rendered figures (preferred for charts, diagrams, chat UIs)

Quickstart for a new paper

1. Copy the starter scaffold:

   cp scripts/render_paper_template.py scripts/render_<paper-slug>.py
   

2. Replace the demo_* functions with gen_figure1(), gen_figure2(), … one per figure in your catalog.
3. For each figure: pick the matching template below, set PX_W/PX_H, plug in your data/labels.
4. Run:

   python3 scripts/render_<paper-slug>.py
   

   Outputs land in output/generated/<timestamp>-<slug>.png.

Shared helpers

All three templates draw from scripts/render_helpers.py. Import what you need:

from render_helpers import (
    DPI,
    BG, AI_BUB, AI_TC, USR_BUB, USR_TC, CODE_BG,
    RED_FG, GRN_FG, GRAY, LGRAY,
    CALL_TEL, FN_YEL, STR_GRN, OLD_RED, NEW_GRN,
    new_canvas, timestamp, save_pixel,
    bubble, alert_line, code_blk, search_blk,
    diagram_box, diagram_drum, uarrow, harrow, arc_arrow, context_brace,
)

Why these helpers exist: matplotlib's FancyBboxPatch expands outward by pad on every side, so the box you draw isn't the box you specified. The rbox() helper inside render_helpers.py compensates for this so you get pixel-exact placement. All other primitives build on top of rbox().

Pixel-based coordinate convention

• DPI = 100, so figsize = (PX_W/DPI, PX_H/DPI) ⟹ 1 data unit = 1 pixel.
• new_canvas(PX_W, PX_H) returns a full-bleed (fig, ax) with no margins; ax.set_xlim(0, PX_W), ax.set_ylim(0, PX_H).
• Y axis goes upward: y=0 is bottom, y=PX_H is top. Stack chat bubbles top-down by decreasing y for each new element.
• Match PX_W/PX_H to your benchmark exactly:

  python3 -c "from PIL import Image; print(Image.open('output/benchmark/figure_X.png').size)"
  

---

Template A — chat_mockup (white-background conversation with code blocks)

Used for figures that show an LLM chatting with a user, often interspersed with system-alert lines and annotated code/search calls. Light-gray AI bubbles with dark text on the left; blue user bubbles with white text on the right; dark navy code/search blocks for syntax-highlighted snippets.

import matplotlib.pyplot as plt
from render_helpers import (
    BG, AI_BUB, AI_TC, USR_BUB, GRAY, RED_FG,
    CALL_TEL, FN_YEL, STR_GRN,
    new_canvas, timestamp, save_pixel,
    bubble, alert_line, code_blk, search_blk,
)

PX_W, PX_H = 320, 170                # ← match target/benchmark dimensions
fig, ax = new_canvas(PX_W, PX_H)

ax.text(PX_W/2, PX_H-10, 'February 7', ha='center', va='center',
        color=GRAY, fontsize=7.5, zorder=3)

# AI (left, light gray + dark text)
bubble(ax, 5, 130, 195, 26, AI_BUB,
       'Assistant message goes here.', tc=AI_TC, fs=7.5)

# User (right, blue + white text — default tc)
bubble(ax, 90, 98, 220, 26, USR_BUB,
       'User reply goes here.', ha='center', fs=7.5)

# Plain-text alert (no surrounding box)
alert_line(ax, 80, 'System Alert: example event', RED_FG, fs=7)

# Single-line code block, syntax-highlighted
code_blk(ax, 5, 50, PX_W-10, 22,
         [[('memory.append(',   FN_YEL),
           ('"example value")', STR_GRN)]], fs=6.5)

# Search call + paginated results
search_blk(ax, 5, 8, PX_W-10, 38,
           call_parts=[('memory.search(', CALL_TEL),
                        ('"query"',        FN_YEL),
                        (')',              CALL_TEL)],
           hdr='Showing 1 of 1 results (page 1/1):',
           results=['"Example result row"'], fs=6.3)

save_pixel(fig, f'output/generated/{timestamp()}-chat.png')

Patterns to remember

• AI bubble: bubble(ax, 5, y, ~195, 26..30, AI_BUB, text, tc=AI_TC).
• User bubble: bubble(ax, ~90, y, ~215, 26..30, USR_BUB, text, ha='center').
• Alert line: alert_line(ax, y, text, RED_FG) for errors/pressure; GRN_FG for success.
• Inline code: code_blk(ax, x, y, w, h, [[(fn, FN_YEL), (str, STR_GRN)]]).
• Multi-line edit (e.g. replace()): pass multiple lines, use OLD_RED/NEW_GRN.
• Search call: call_parts = [(fn, CALL_TEL), (term, FN_YEL), (')', CALL_TEL)].
• Line height in code/search blocks: (h - 8) / n_lines ≈ 11–13 px at fs=6.5.

---

Template B — architecture_diagram (colored boxes with arrows)

Used for system architecture figures with named components, data-flow arrows, optional storage cylinders, and labeled tiers.

import matplotlib.pyplot as plt
from render_helpers import (
    DPI, timestamp, save_pixel,
    diagram_box, diagram_drum,
    uarrow, harrow, arc_arrow, context_brace,
)

PX_W, PX_H = 800, 320
fig = plt.figure(figsize=(PX_W/DPI, PX_H/DPI))
fig.patch.set_facecolor('white')
ax = fig.add_axes([0, 0, 1, 1])
ax.set_xlim(0, PX_W);  ax.set_ylim(0, PX_H);  ax.axis('off')

# Optional grouping label (overbrace)
context_brace(ax, 30, 540, 290, 'Logical Group Label', fs=12)

# Component boxes — fc/tc/fw/fs are per-component
diagram_box(ax,  30, 160, 220, 110, 'Component\nA',
            fc='#1a1a1a', tc='white', fw='bold', fs=13)
diagram_box(ax, 280, 160, 220, 110, 'Component\nB',
            fc='#f08c00', tc='#111', fw='bold', fs=13)
diagram_box(ax, 530, 160, 220, 110, 'Component\nC',
            fc='white',   tc='#111', fw='bold', fs=13, dashed=True)

# Storage cylinder
diagram_drum(ax, 30, 20, 180, 110, 'Storage', fs=13)

# Arrows: choose by type
uarrow(ax, 390, 130, 160, '#f08c00', lw=3.0)        # vertical UP
harrow(ax, 250, 280, 215, '#333')                    # horizontal →
arc_arrow(ax, 730, 215, 250, 215, '#4080c0',         # curved ↻
          rad=-0.4)

# Optional dashed bounding rectangle around a logical group
ax.add_patch(plt.Rectangle((25, 155), 480, 120,
    fill=False, ec='black', lw=1.3, ls='--', zorder=1))

save_pixel(fig, f'output/generated/{timestamp()}-arch.png', bg='white')

Patterns to remember

• Box colors: each component gets (fc, tc, fw). Pick a consistent palette per paper — fill colors usually carry semantic meaning (e.g. read-only vs read-write).
• Drum (storage cylinder): diagram_drum() for any persistent store / database.
• Vertical UP arrow: uarrow(ax, x, y_bottom, y_top, color, lw).
• Horizontal arrow: harrow(ax, x0, x1, y, color). For bidirectional, draw two parallel arrows offset by ~10–20 px.
• Curved arc: arc_arrow(...) — negative rad curves below the straight line, positive above. Useful for long-range connections that would cross other elements.
• Dashed grouping rectangle: plt.Rectangle((x, y), w, h, fill=False, ls='--', ec='black', lw=1.3) for "this set of boxes belongs together" semantics.
• Top brace + label: context_brace() gives the LaTeX-\overbrace look common in systems papers.

---

Template C — line_chart (accuracy vs X-axis, multiple models)

Used for performance result figures comparing multiple models across a scalar axis (documents retrieved, nesting level, epoch, etc.).

Critical — extract data from the paper text first.

1. Tables: exact accuracy/F1/ROUGE numbers usually live in tables — copy them verbatim.
2. Results prose: the paper's narrative tells you the shape of each curve (monotonic? rises-then-falls? plateaus?). Always reread it before plotting.
3. Visual estimation from a benchmark image: last resort only — small charts make it easy to mis-read by ±0.05 absolute accuracy.

Real-world example: in MemGPT Figure 5, fixed-context baselines (GPT-4, GPT-3.5) RISE then FALL as documents fill the context window — not monotonically decrease. Reading the paper text avoids this kind of mistake.

import matplotlib.pyplot as plt
from render_helpers import DPI, timestamp

X = [0, 1, 2, 3, 4]
SERIES = [
    # (label,         ys,                              color,    ls,    marker, ms)
    ('Baseline A',    [0.40, 0.55, 0.65, 0.50, 0.30], '#1f77b4', '--', 's', 4),
    ('Baseline B',    [0.35, 0.45, 0.55, 0.40, 0.25], '#aec7e8', '-.', 'D', 4),
    ('Our Method',    [0.65, 0.68, 0.70, 0.69, 0.66], '#2ca02c', '-',  'o', 5),
]

fig, ax = plt.subplots(figsize=(3.5, 2.3))
fig.patch.set_facecolor('white');  ax.set_facecolor('white')
for label, ys, color, ls, marker, ms in SERIES:
    ax.plot(X, ys, color=color, linestyle=ls, marker=marker,
            markersize=ms, linewidth=1.2, label=label)
ax.set_xlabel('X-axis label', fontsize=8)
ax.set_ylabel('Accuracy',     fontsize=8)
ax.set_ylim(0.0, 0.8)
ax.grid(True, linestyle='-', linewidth=0.3, color='#e0e0e0', zorder=0)
ax.tick_params(labelsize=7)
ax.legend(fontsize=6, loc='lower left', framealpha=0.95,
          handlelength=2.2, handletextpad=0.5)
for spine in ax.spines.values():
    spine.set_linewidth(0.6);  spine.set_color('#bbb')
plt.tight_layout(pad=0.3)
plt.savefig(f'output/generated/{timestamp()}-chart.png',
            dpi=DPI, facecolor='white')
plt.close(fig)

Patterns to remember

• Color palette: stay close to matplotlib's defaults (#1f77b4, #ff7f0e, #2ca02c, …) — papers tend to use these.
• Line style mapping: solid for "ours" / hero method; dashed/dotted for baselines. Distinct markers also help when colors print B&W.
• Y-limits: trim aggressively. Don't waste vertical space on (0, 1) when results live in (0.05, 0.75).
• Tiny figures: when matching a benchmark thumbnail (e.g. 320×210 px), use figsize=(PX_W/100, PX_H/100) and shrink fontsize to 6–8.
• Bar charts: same approach with ax.bar() + hatch patterns instead of markers.

---

Step 4b — Nano Banana JSON prompt (for conceptual / organic figures)

Use this path only for schematic_flow, conceptual, venn, and quadrant figures that are difficult to express in code.

Read these references before drafting:

• references/nano-banana-prompt-shape.md — canonical JSON skeleton.
• references/paper-figure-style.md — default LaTeX-style cues.
• references/figure-types.md — figure-type cheat sheet.
• references/whiteboard-style.md — alternate hand-drawn template.

Default aesthetic: publication-ready paper figure. Clean black-on-white schematic, rounded rectangle nodes, thin strokes, labeled arrows, Computer Modern Serif typography.

Text-rendering reality check. Diffusion models mangle long labels and equations:

• Keep every label 1–4 real English words.
• At most one boxed equation per figure.
• Re-render 2–4 times; tweak the offending label as a separate weighted_positive line.

Required top-level shape:

{
  "pipeline_configuration": {
    "job_type": "txt2img_generation",
    "meta_tags": ["..."],
    "generative_parameters": { },
    "text_prompts": {
      "weighted_positive": { "...": 1.3 },
      "weighted_negative": { "...": 1.4 }
    }
  }
}

Save and validate:

cat > prompt.json <<'JSON'
{ "pipeline_configuration": { ... } }
JSON
python3 -c "import json; json.load(open('prompt.json')); print('valid')"

✋ Stop here unless json.load returns valid.

---

Step 5 — Generate the image

Code-path figures

python3 scripts/render_<paper-slug>.py

The script saves directly to output/generated/<timestamp>-<slug>.png. No additional step needed.

Nano Banana figures

Draft (txt2img):

python3 scripts/generate.py \
  --prompt "$(cat prompt.json)" \
  --filename "output/generated/$(date +%Y-%m-%d-%H-%M-%S)-paper-draft.png" \
  --resolution 1K

✋ Stop here unless the 1K draft is on-target. Only escalate to 4K once the draft is right.

Final (4K):

python3 scripts/generate.py \
  --prompt "$(cat prompt.json)" \
  --filename "output/generated/$(date +%Y-%m-%d-%H-%M-%S)-paper-final.png" \
  --resolution 4K

Img2img (when a reference image is provided):

python3 scripts/generate.py \
  --prompt "$(cat prompt.json)" \
  --filename "output/generated/$(date +%Y-%m-%d-%H-%M-%S)-paper-edit.png" \
  --input-image path/to/reference.png \
  --resolution 2K

Always run from the project root so output/ resolves correctly.

---

Calibration loop (when a benchmark exists)

If the user supplies reference images in output/benchmark/, treat them as the visual target:

1. Match PX_W/PX_H exactly with Image.open(...).size.
2. Generate the figure.
3. Open both images side-by-side. Compare:
   • Element positions (AI bubble height, code block width, …).
   • Color values (especially bubble fills, code highlight colors).
   • Text content (typos, formatting, line breaks).
   • Arrow targets (which box does each arrow connect to?).
   • Chart data (does each series have the right shape per the paper text?).
4. Fix one thing at a time, regenerate, re-compare.

The MemGPT reference (scripts/render_all_memgpt.py ↔ output/benchmark/figure_*.png) shows what "matches benchmark" looks like in practice.

---

Worked reference example — MemGPT paper

The MemGPT paper ("Towards LLMs as Operating Systems", arXiv 2310.08560) demonstrates all three figure archetypes. Use this as a calibration target — once you can reproduce its 8 figures from output/benchmark/, the same primitives work for any other paper.

Paper in brief

MemGPT treats the LLM as an OS: it moves data between a fixed-size main context (the LLM's context window) and unbounded external storage (recall + archival). A Function Executor and Queue Manager handle paging, similar to virtual memory.

Figure catalog (MemGPT)

#	File	Description	Type	Render
1	output/benchmark/figure.png	Birthday chat, memory-pressure alert, working_context.append()	chat_mockup	code
2	output/benchmark/figure_2.png	"Six flags" recall search with paginated results	chat_mockup	code
3	output/benchmark/figure_3.png	Full system architecture: context window tiers + external storage	architecture_diagram	code
4	output/benchmark/figure_4.png	Feb 14 breakup, working_context.replace()	chat_mockup	code
5	output/benchmark/figure_5.png	DMR accuracy vs documents retrieved (5 model lines)	line_chart	code
6	output/benchmark/figure_6.png	Nobel Prize archival storage search, paginated	chat_mockup	code
7	output/benchmark/figure_7.png	Multi-session chat accuracy vs nesting level	line_chart	code
8	output/benchmark/figure_8.png	Key-value store search through archival storage	chat_mockup	code

Reproduce all 8 figures with:

python3 scripts/render_all_memgpt.py

Chat mockup colour values (verified against output/benchmark/)

Chat figures use a white page background. Code/search blocks retain a dark navy background for contrast.

Element	Hex	Notes
Page background	#ffffff	White
AI bubble (left)	#e5e5ea	iOS light gray; pair with tc=AI_TC
AI bubble text	#1c1c1e	Dark, readable on light gray
User bubble (right)	#1a6bcd	Blue; pair with white text
Code/search block background	#1a1a2e	Dark navy
System alert text (memory)	#ff3b30	Plain text, NO box
System alert text (archive done)	#30d158	Plain text, NO box
Function name in search calls	#4ec9b0	Teal CALL_TEL
Search term / string literals	#dcdcaa	Yellow FN_YEL
Code block string literals	#4ec9b0	Teal STR_GRN
Old value in replace()	#f44747
New value in replace()	#4ec9b0
Result body text	#c8c8d8
Date header	#8888aa

Architecture diagram (Figure 3) — verified arrow pattern

• Orange UP: Function Executor → Working Context.
• Pink UP: Queue Manager → FIFO Queue (left position).
• Blue UP: Queue Manager → FIFO Queue (right position, slight offset). Two arrows to FIFO Queue.
• Green bidirectional (horizontal): Archival Storage ↔ Function Executor.
• Black bidirectional (horizontal): Function Executor ↔ Queue Manager.
• Blue ONE-WAY right (horizontal): Queue Manager → Recall Storage (Write via Queue Manager).
• Large blue arc (curved, below): Recall Storage → Function Executor (Read via Functions).
• Dashed bounding rectangle: around the 3 Prompt-Tokens boxes (System Instructions + Working Context + FIFO Queue), NOT around Output Buffer. Output Buffer has its own separate dashed border.

Line charts (Figures 5 & 7) — values

These were extracted from the paper's tables and prose, then cross-checked against the benchmark image:

• Figure 5 (Document QA): fixed-context models RISE then FALL as documents fill the 8k window; MemGPT (GPT-4 / GPT-4 Turbo) stays flat ~0.67. GPT-4 series: [0.40, 0.60, 0.70, 0.60, 0.38, 0.28, 0.20, 0.14, 0.10] over X = [0, 25, 50, 75, 100, 125, 150, 175, 200].
• Figure 7 (Nested KV): MemGPT (GPT-4 Turbo) plateaus at 1.0 for levels 0–1 then degrades; baselines drop to ~0 by level 3. The paper notes that "MemGPT with GPT-4 Turbo performs worse than MemGPT with GPT-4", so MemGPT (GPT-4) holds up better at higher nesting.

Always cross-check chart shapes against the paper's Results section before committing data values.

---

API keys

Keys are read from environment variables — never stored in any committed file.

Provider	Env var	Get key at
Gemini	GEMINI_API_KEY	https://aistudio.google.com/apikey
fal.ai	FAL_KEY	https://fal.ai/dashboard/keys

Add to your shell profile:

export GEMINI_API_KEY=<your-key>
export FAL_KEY=<your-key>

If the user pasted a key in chat, treat it as compromised — ask them to rotate it.

See references/providers.md for full setup, model options, and common error fixes.

---

Iteration tips

• Keep Python render scripts; tweak one parameter at a time between runs.
• One change per iteration so cause and effect are clear.
• For code-path figures: each run produces a new timestamped file — no overwrites.
• For Nano Banana figures: if a label renders garbled, simplify it or move it to the caption.
• When in doubt about a chart shape, reread the paper's results section before tweaking pixels.

Resources

• scripts/render_helpers.py — palette + drawing primitives (import these into every render script).
• scripts/render_paper_template.py — starter scaffold for a new paper.
• scripts/render_all_memgpt.py — worked reference example (all 8 MemGPT figures).
• scripts/setup.py — interactive provider setup (Nano Banana only).
• scripts/generate.py — unified Nano Banana dispatcher.
• references/providers.md — provider setup, API keys, model lists.
• references/nano-banana-prompt-shape.md — JSON skeleton.
• references/paper-figure-style.md — LaTeX paper-figure aesthetic.
• references/whiteboard-style.md — whiteboard-sketch aesthetic.
• references/figure-types.md — figure-type cheat sheet.
• references/img2img-preservation.md — img2img preservation rules.