Lora Visual

Lora Visual

Create raster visuals in one shared cute Japanese-anime editorial language. Choose one initial output mode before generating. Do not run both production paths in parallel. Mode A may switch once to Mode B only after reaching the text-repair stopping condition below.

Use assets/lora-character-reference-v2.png as the canonical character reference whenever the image tool supports references. It is deliberately free of text, logos, labels, and pseudo-text.

Choose the output mode

Mode A — explanatory image

Use when the image must explain a concept, mechanism, workflow, comparison, or tradeoff by itself.

• Deliver a complete PNG or WebP with a finished off-white scene.
• Generate every essential title and label directly inside the bitmap.
• Make the relation visible through objects, paths, states, or repeated materials; labels identify the evidence but do not replace it.
• Do not generate an unlabeled base and add essential words in a separate rendering step.

Mode B — transparent illustration

Use when the character scene will be composed into a hero, document, card, slide, or other layout.

• Generate the subject on a perfectly uniform chroma-key background that does not occur in the artwork. Default to #00FF00; use #FF00FF when the subject contains green.
• Do not include explanatory labels unless the user explicitly requests text inside the illustration.
• Validate the keyed source with scripts/validate_mode_b.py before removal.
• Remove the background with scripts/cutout.py, validate the transparent result, and deliver a transparent PNG.
• Keep the transparent artwork as a reusable visual asset; the surrounding layout supplies the title and explanatory copy.

If the destination is unclear, choose Mode A when the image itself must communicate the idea and Mode B when another layout will carry the explanation.

Shared visual language

• Keep the recurring characters visually consistent with assets/lora-character-reference-v2.png.
• lora is a cute young-adult Japanese-anime heroine with dark brown-black hair in a loose high messy bun, soft face-framing strands, large warm dark-brown eyes, a small blank amber hair clip with no letters or symbols when visible, an oversized cream hoodie, loose dark cargo pants, and black-and-white sneakers.
• Mochi is a small fictional warm-cream fluffy companion dog with floppy ears, round dark eyes, a black nose, and a compact friendly silhouette.
• Treat lora and Mochi as a paired cast. Every newly generated image that includes lora must also include Mochi in the same scene or transparent asset. Omit Mochi only when the user explicitly requests lora alone or explicitly says not to include Mochi. Limited space, composition, props, or mode choice are not valid reasons to omit Mochi.
• Keep characters secondary to the subject's evidence or action.
• Use black, off-white, cream, warm gray, brown, and amber as the base visual system. Amber is the main identity accent.
• Add at most one additional muted semantic color when the explanation truly needs it.
• Favor clean anime line work, restrained paper texture, notebook/editorial materiality, and a warm creator-world atmosphere.
• Avoid 3D, glossy gradients, photorealism, purple-blue AI gradients, neon cyberpunk, generic card grids, dashboards, decorative clutter, excessive pink, and watermarks.

Mode A workflow — explanatory image

1. Write the visual brief

viewer_question: what should be understood in 10 seconds?
concrete_claim: one-sentence conclusion
real_objects: visible objects, interfaces, documents, tools, or states
relation: comparison, transformation, causality, sequence, hierarchy, feedback, tradeoff, or pipeline
visual_evidence: what must remain understandable when labels are ignored?
scene: believable setting and 2–4 useful environmental cues
semantic_colors: what each accent color means
labels: exact short strings plus the evidence surface for each label

Show the input, action or relation, and result. Keep one dominant focal action and no more than three major visual regions. For multiple steps, use a simple left-to-right or top-to-bottom sequence.

2. Design the labels

• Prefer 2–6 labels. Use more only when the explanation truly needs them.
• Keep each label short and concrete: role, action, state, or outcome.
• Place every label on or immediately beside its evidence surface, such as a desk nameplate, task sheet, folder tab, machine, meter, lane, or result document.
• Use modern Chinese sans-serif typography, medium or bold, large enough to read at the intended display size.
• Do not turn body copy, commands, tables, or long paragraphs into image text. Use a deterministic layout method when dense or editable text is required.

Add this block to the generation prompt:

Text (verbatim): Render these exact labels as part of the bitmap illustration:
"<label 1>", "<label 2>", "<label 3>".
Use each phrase exactly once. Do not translate, paraphrase, misspell, repeat,
or add any other text. Use modern sans-serif medium/bold typography, large
and readable. Place "<label 1>" on <evidence surface>; place "<label 2>" on
<evidence surface>; place "<label 3>" on <evidence surface>.

3. Build the prompt

Use this order:

1. State the concrete claim and shared task.
2. Describe the real setting and the lora/Mochi action.
3. Describe the evidence objects and their geometry: aligned, nested, connected, split, transformed, repeated, or converging.
4. Assign semantic colors.
5. Quote the exact labels and specify each placement.
6. Add the Mode A style anchor.
7. End with exclusions.

Mode A style anchor:

Cute Japanese-anime editorial illustration with clean confident charcoal/black
line work and restrained warm paper texture. Use the canonical lora reference:
dark brown-black loose high messy bun, soft face-framing strands, large warm
dark-brown eyes, small blank amber hair clip with no letters or symbols when
visible, oversized cream hoodie, loose dark cargo pants, and black-and-white
sneakers. Mochi must appear in the same scene unless the user explicitly asks
for lora alone or says not to include Mochi. Mochi is a small warm-cream fluffy
companion dog with floppy ears, round dark eyes, a black nose, and a compact
friendly silhouette. Do not omit Mochi to simplify the composition. Use an
off-white lightly textured real environment, not a blank white canvas.
Typography is modern sans-serif, medium or bold, large and readable.
Color is restrained: black, off-white, cream, warm gray, brown, and amber,
plus at most one muted semantic accent color. Keep the scene cute, thoughtful,
clean, warm, and editorial. No 3D, no glossy gradients, no photorealism,
no purple-blue AI gradient, no neon cyberpunk, no generic card grid,
no dashboard, no decorative clutter, no excessive pink, no tiny text,
no long paragraphs, no watermark.

4. Inspect, retry, and stop

1. Inspect the output at original resolution.
2. Compare every label with the brief character by character. Confirm that each appears exactly once and that no stray text was added.
3. Reject missing, duplicated, invented, or misspelled labels.
4. Make at most two targeted text-repair attempts after the initial generation. Repeat all scene and style invariants on each attempt. Do not conceal an error with a separate text layer.
5. If any required text is still wrong after the second repair, stop bitmap text generation. Do not make a third repair.
6. Switch once to Mode B for the same claim and composition. Generate a no-text transparent illustration, then place the exact title and labels with deterministic SVG typography. Preserve the visual evidence, reading order, palette, and intended display size. Report that the Mode A stopping condition triggered the fallback.

Use this retry instruction:

Keep the scene, composition, characters, objects, colors, and all correct labels unchanged.
Change only the incorrect text "<wrong>" to the exact text "<right>".
Do not add, remove, translate, or repeat any other text.

The Mode B plus SVG fallback is the terminal path for that request. If the transparent illustration or SVG composition fails its own validation, stop and report the failed check instead of returning to Mode A.

Mode B workflow — transparent illustration

1. Describe one reusable scene

Use one paired lora-and-Mochi action and only the objects needed to establish it. Examples: lora thinking while Mochi watches, lora writing with Mochi beside the notebook, lora pointing while Mochi follows the gesture, or both celebrating a finished result. Leave generous padding around the complete pair so the cutout can be composed safely. If the user explicitly requests lora alone or no Mochi, follow that request.

2. Build the prompt

Describe the subject first, then append this fixed anchor:

Replace <KEY_COLOR> with the selected hex color before sending the prompt.

Style: cute Japanese-anime editorial character illustration. Clean confident
charcoal/black line work with controlled detail, not photorealistic and not 3D.
Use the canonical lora reference: dark brown-black loose high messy bun, soft
face-framing strands, large warm dark-brown eyes, small blank amber hair clip
with no letters or symbols when visible, oversized cream hoodie, loose dark cargo
pants, and black-and-white sneakers. Mochi must appear in the same transparent
asset unless the user explicitly asks for lora alone or says not to include
Mochi. Mochi is a small warm-cream fluffy dog with floppy ears, round dark eyes,
a black nose, and a compact friendly silhouette. Keep both characters fully
visible with safe space between them and the image border. Color is restrained:
black, off-white, cream, warm gray, brown, and amber.

The background must be a perfectly uniform flat <KEY_COLOR> rectangle with zero
gradient, texture, noise, speckles, shadows, floor plane, or lighting variation.
Do not let line work, props, hair, shoes, dog ears, tail, or the subject touch
the image border. Keep generous padding. No text, no watermark. PNG format.

3. Validate the source and repair the background once

• Inspect the image before removal.
• Confirm all four corners are uniform and visually match the chosen key color.
• Reject backgrounds with gradients, texture, shadows, speckles, or artwork touching the border.
• Preserve the source image alongside the transparent result until the output is approved.

Run the automatic source checks before background removal:

python3 scripts/validate_mode_b.py source.png --expected-key '#00FF00'

Replace the expected key when using magenta. The validator checks border color variance, distance from the expected key, and whether the subject enters the required edge margin.

If the first source fails only source.background_uniform or source.expected_key while source.subject_padding passes, make exactly one background-only edit. Do not regenerate or restyle the character.

Keep the character, face, hair, blank amber hair clip, expression, pose, clothing,
props, Mochi, composition, crop, scale, line work, colors, and all subject pixels
unchanged. Change only the background to one perfectly uniform flat <KEY_COLOR>
rectangle. Remove every background gradient, texture, shadow, floor, reflection,
speckle, and lighting variation. Do not redraw, move, crop, or restyle the subject.
No text, no letters, no symbols, no logo, no watermark.

Validate the edited source again. If it still fails, or if the subject touches the edge, stop and report the failed checks. Do not keep editing the background and do not send an invalid source to cutout.py.

4. Remove the background

Install Pillow if the active Python environment does not have it, then run:

python3 scripts/cutout.py source.png transparent.png

Optional tuning:

python3 scripts/cutout.py source.png transparent.png \
  --transparent-threshold 12 \
  --opaque-threshold 220

The script samples the image border, builds a soft alpha matte from color distance, and removes color spill from antialiased edges. It works with any uniform key color, so the key can be chosen to avoid the subject palette.

5. Validate the transparent result

Run the combined source and transparent checks:

python3 scripts/validate_mode_b.py \
  source.png transparent.png \
  --expected-key '#00FF00'

The validator requires an RGBA output, four transparent corners, no key-color fringe above the allowed ratio, and subject padding on every edge. It exits nonzero on failure and supports --json for automation.

• Confirm the output is RGBA and all four corners have alpha 0.
• Confirm the subject remains complete, including loose hair strands, hoodie strings, shoes, the blank amber hair clip, Mochi's ears and tail, and small props.
• Check for a gray, green, or magenta fringe at 100% zoom.
• Confirm internal cream / white areas were not erased.
• If only the transparent result fails, tune cutout.py once without regenerating the character, then validate again. Stop and report if the second transparent result still fails.

Regression references

Read evals/EVALS.md only when testing or updating this skill. The images under assets/golden/ are regression references, not content to insert into ordinary outputs. See manifest.json for versioned asset hashes and upstream provenance.

Output handling

• Save approved project assets inside the current project or output directory.
• Do not leave project-referenced images only in the generator's default storage.
• Use versioned filenames instead of overwriting an approved asset unless the user explicitly requests replacement.
• Report the final prompt, output mode, source image path when Mode B is used, final image path, validator result, fallback reason when Mode A switches to Mode B, and any non-default cutout options.

Quality gate

For every output:

• The subject is recognizable in about 3 seconds.
• The main action or relation is clear in about 10 seconds.
• Characters support the subject instead of becoming generic decoration.
• lora, Mochi, the black/cream/amber palette, and the warm editorial feel remain consistent.
• Every output containing lora also contains a recognizable, complete Mochi unless the user explicitly requested lora alone or no Mochi.

For Mode A:

• One claim, one focal action, and no more than three major visual regions.
• The visual evidence still shows the relation when labels are ignored.
• Every required label is exact, appears once, and is integrated into the correct evidence surface.
• All labels remain readable at the intended display size.
• No request exceeds two targeted text-repair attempts. A still-invalid result switches once to Mode B plus deterministic SVG typography.

For Mode B:

• Background removal is clean and the output has real transparency.
• Thin details and internal cream/white regions remain intact.
• No source background, fringe, shadow, or border artifact remains.
• scripts/validate_mode_b.py passes for the keyed source and transparent output.