Fix T2I sheets: four-panel photorealistic turnaround instead of design board

- references/t2i-character-sheet.md: rewrite around the fixed four-panel layout
  (face close-up, front, side 90deg right, back) on a white background with no
  text/labels/borders; add photorealism rules and banned illustration vocabulary
- worked-example.md: regenerate Image-1.json / Image-2.json in the new style
- validate_outputs.py: guard sheet captions (four-panel structure, photorealistic
  language, banned cartoon/design-board vocabulary)
- SKILL.md: step 6 + checklist item 9 updated accordingly
This commit is contained in:
Z User 2026-11-11 22:50:14 +00:00
parent 8759ead141
commit 9e931a5c7c
4 changed files with 95 additions and 52 deletions

View File

@ -74,8 +74,11 @@ When a subject sings to the provided audio, use the exact fixed audio template a
exact lip-sync phrase defined in that reference. exact lip-sync phrase defined in that reference.
**Step 6 — Write the character sheets** per `references/t2i-character-sheet.md`, one **Step 6 — Write the character sheets** per `references/t2i-character-sheet.md`, one
`Image-N.json` per subject lacking a user reference. Captions must reuse the locked `Image-N.json` per subject lacking a user reference. Every caption describes the
story-bible values verbatim. fixed four-panel photorealistic turnaround (face close-up, front, side facing 90°
right, back) on a white background with no text, no labels, and no borders — studio
photography of a real person, never a cartoon or designed card — and must reuse the
locked story-bible values verbatim.
**Step 7 — Write the manifest** `image-map.txt`, one line per subject: **Step 7 — Write the manifest** `image-map.txt`, one line per subject:
@ -133,11 +136,17 @@ folder.
`detailed_description`. Non-singing segments use `[reference generation]`. `detailed_description`. Non-singing segments use `[reference generation]`.
7. `[Shot 1]` has no timestamp; any `At MM:SS.mmm` appears only on `[Shot N]`, N≥2. 7. `[Shot 1]` has no timestamp; any `At MM:SS.mmm` appears only on `[Shot N]`, N≥2.
8. Sheet captions and segment descriptions agree with the locked story bible. 8. Sheet captions and segment descriptions agree with the locked story bible.
9. Each sheet caption describes the four-panel photorealistic turnaround (face
close-up panel, full-body front, side facing 90° right, back) on a clean white
background with equal spacing and no text, no labels, no borders; it uses
photorealistic/photographic language and contains no cartoon, illustration, or
design-board vocabulary (cartoon, anime, mascot, name tag, swatch, palette strip,
design board, CGI, 3D render...).
## References ## References
- `references/ref2va-segment-format.md` — authoritative per-segment format (sections, - `references/ref2va-segment-format.md` — authoritative per-segment format (sections,
labels, shots, audio template, retention markers). labels, shots, audio template, retention markers).
- `references/t2i-character-sheet.md` — character-sheet caption rules and the two-key - `references/t2i-character-sheet.md` — the four-panel photorealistic turnaround
JSON output contract. sheet rules and the two-key JSON output contract.
- `references/worked-example.md` — full end-to-end example, including an audio variant. - `references/worked-example.md` — full end-to-end example, including an audio variant.

View File

@ -1,8 +1,10 @@
# T2I Character Sheet Prompts (Image-N.json) # T2I Character Sheet Prompts (Image-N.json)
This reference defines how the skill writes one character-sheet text-to-image prompt This reference defines how the skill writes one character-sheet text-to-image prompt
for a character that has no user-provided image reference. It adapts the caption- for a character that has no user-provided image reference. The JSON output contract
expansion and JSON contract rules in `t2i-prompt-rules/` of this repo. comes from `t2i-prompt-rules/` of this repo; the sheet itself is a fixed FOUR-PANEL
PHOTOREALISTIC TURNAROUND on a white background. It must read as studio reference
photography of a real person — never a cartoon, illustration, or designed card.
## 1. Output contract (exact) ## 1. Output contract (exact)
@ -18,8 +20,10 @@ fence, no commentary, no extra keys:
general ratio-selection policy in the source rules. general ratio-selection policy in the source rules.
- `expanded_caption` is ONE fluent English paragraph (a string). Every ASCII double - `expanded_caption` is ONE fluent English paragraph (a string). Every ASCII double
quote used inside the caption is serialized as the JSON escape `\"`; escape all quote used inside the caption is serialized as the JSON escape `\"`; escape all
other characters as valid JSON requires. other characters as valid JSON requires. Prefer wording without quotes at all.
- The file must parse with a standard JSON parser. - The file must parse with a standard JSON parser.
- Write the file in the compact single-line form shown in the example: one JSON
object, two keys, no surrounding text.
## 2. What the sheet is for ## 2. What the sheet is for
@ -27,56 +31,73 @@ The rendered sheet becomes the character's reference image: it is cited as
`<Picture N>` in the segment prompts and drives the character's identity in every `<Picture N>` in the segment prompts and drives the character's identity in every
segment. Therefore the caption must make the character's canonical look fully segment. Therefore the caption must make the character's canonical look fully
inspectable, and its every detail must match the locked story bible used by the inspectable, and its every detail must match the locked story bible used by the
segment prompts — same hair, eyes, build, wardrobe colors, materials, props. If the segment prompts — same hair, eyes, skin tone, build, wardrobe colors and patterns,
vague prompt left a trait unspecified, the skill invents it once (Step 3 of the materials, footwear, worn props. If the vague prompt left a trait unspecified, the
pipeline) and uses the identical wording here and in the segments. skill invents it once (Step 3 of the pipeline) and uses the identical wording here
and in the segments.
## 3. Caption design rules ## 3. Required sheet structure (fixed four-panel turnaround)
Act as an expert visual designer. The requested image is a communication artifact — Every caption describes EXACTLY this layout, on a clean white background:
a designed character card / model sheet on a horizontal 16:9 canvas — not a plain
portrait and not a scene. Build a finished information architecture:
Required elements (weave them into one flowing caption paragraph): 1. **Frame**: a photorealistic full-color character reference sheet on a clean white
background, divided into FOUR panels arranged side by side, with equal white
spacing between the panels, and no text, no labels, no borders, no watermark.
2. **Panel 1 (leftmost, narrower panel)**: a CLOSE-UP of the person's FACE ONLY —
head and shoulders, front-facing, filling most of the panel. The face must be
fully visible: helmets, hoods, hats, and masks are off or carried, never covering
the face in this panel.
3. **Panel 2**: FULL BODY from head to toe, FRONT view, facing directly forward,
no cropping.
4. **Panel 3**: FULL BODY from head to toe, SIDE view, facing 90 degrees to the
right, no cropping.
5. **Panel 4**: FULL BODY from head to toe, BACK view, facing completely away —
back of the head and body visible, no cropping.
6. **Identity lock**: the SAME person in all four panels — identical facial
features, skin tone, hair color, and hairstyle, and the EXACT same clothing
(same garments, colors, and patterns) in every panel.
1. **Frame the artifact**: state that this is a 16:9 horizontal character design ## 4. Photorealism rules (never cartoons)
board / model sheet for the named character, clean designed layout.
2. **Identity block**: title/name tag with the character's name (invent a fitting
name if the prompt gives none), role line, optional short profile/about note.
3. **One large full-body view**: the character standing in a neutral, readable
stance (front or slight three-quarter), full figure with readable proportions,
feet visible, no cropping.
4. **Inspectable trait detail**: hair style and color, eye color, skin tone, build,
each wardrobe piece with its color and material, footwear, and any signature
props — exactly the locked story-bible values.
5. **Attribute rows / callouts**: labeled swatch or annotation elements for key
colors (palette strip) and materials, so the board reads as a designed sheet.
6. **Props tied to the role** displayed beside or held by the figure when the role
implies them.
7. **Optional supporting elements** that add completeness without clutter: a short
quote or badge, a small contextual backdrop hint, an instruction row. Keep these
as short readable on-image labels, not prose blocks.
Hard rules carried over from the source rules: The sheet always depicts a real human being captured photographically:
- Preserve every hard fact from the user's prompt (names, stated traits, counts). - Positive style cues to include: photorealistic, full-color photograph, studio
- Required content must be visible, not implied; if a label is described, say where reference photography, natural skin texture, realistic eyes and hair, true-to-life
it sits on the board. fabric texture and drape, even soft studio lighting, sharp focus, realistic
- One character per sheet. No other characters; background figures never compete. proportions.
- Replace abstract mood and generic adjectives ("cool", "epic") with specific - NEVER use vocabulary that pulls the generation toward illustration or graphic
visible cues (colors, materials, light). design — this is what produces cartoons. Banned words in the caption include:
- The caption is a description of the image, not instructions to the viewer; write cartoon, anime, manga, illustration, stylized, CGI, 3D render, mascot, avatar,
it as one descriptive paragraph in present tense. character design board, name tag, swatch, palette strip, drafting paper. Also do
not describe any text rendered in the image (no titles, tags, or captions on the
sheet itself — the layout rule already forbids text).
- Do not describe background scenery, props floating as insets, or decorative
layouts. The only content is the four panels and the white background.
## 4. Example For non-human characters (creatures, robots) keep the identical four-panel
turnaround: Panel 1 is the close-up of the face or front, and the materials are
described photorealistically (real fur, worn metal, rubber) rather than as toy or
figure renders.
## 5. Filling in the subject
- The skill never has an input image for a generated sheet, so the caption itself
must carry the full likeness: age and gender, build, hair color and style, eye
color, skin tone, each garment (color, pattern, material), footwear, and
worn or carried props — verbatim from the locked story bible.
- Worn props appear wherever physically visible (e.g. a wrist scanner on the arm in
all panels); carried items (e.g. a helmet under the arm) appear in the full-body
panels only, never covering the face in Panel 1.
- One subject per sheet; no other people anywhere in the image.
- Preserve every hard fact from the user's prompt (names, stated traits, counts);
write the caption as one descriptive paragraph in present tense.
## 6. Example
For an astronaut character locked as: woman in her thirties, short black hair, brown For an astronaut character locked as: woman in her thirties, short black hair, brown
eyes, matte-white EVA suit with orange stripe accents, wrist-mounted scanner, clear eyes, fair skin, matte-white EVA suit with orange stripe accents, grey reinforced
dome helmet: gloves, wrist-mounted scanner on the left forearm, clear dome helmet:
```json ```json
{"expanded_caption": "A 16:9 horizontal character design board titled \"VOSS\" in a clean technical sans-serif name tag, presenting a female astronaut in her thirties with short black hair and brown eyes, shown in one large full-body view standing in a neutral stance, feet fully visible. She wears a matte-white EVA suit with orange stripe accents along the arms and chest, a clear dome helmet, grey reinforced gloves, and a wrist-mounted scanner with a small amber readout on her left forearm. A role line under the name tag reads \"Deep-Field Survey Specialist\", and a short profile note describes a calm, methodical explorer. Along the bottom edge, a labeled palette strip shows matte white, safety orange, graphite grey, and amber swatches, with thin callout lines pointing to the suit fabric, helmet glass, and scanner housing. Two signature props, the wrist scanner and a compact sample canister, are displayed as small inset items beside the figure on a light grey drafting-paper background with subtle grid lines, so the board reads as a designed character sheet rather than a plain portrait.", "aspect_ratio": "16:9"} {"expanded_caption": "A photorealistic full-color character reference sheet photographed on a clean white background, divided into FOUR panels arranged side by side with equal white spacing between the panels, and no text, no labels, and no borders. Panel 1, the leftmost and narrower panel, is a close-up of the person's face only — head and shoulders, front-facing, filling most of the panel: a woman in her thirties with short black hair, brown eyes, fair skin, and natural skin texture. Panels 2 through 4 show the same woman full body from head to toe with no cropping: Panel 2 a front view facing directly forward, Panel 3 a side view facing 90 degrees to the right, and Panel 4 a back view facing completely away with the back of her head and body visible. All four panels show the identical person with the same facial features, skin tone, and hairstyle, wearing the exact same clothing: a matte-white EVA suit with orange stripe accents along the arms and chest, grey reinforced gloves, a wrist-mounted scanner with a small amber readout on her left forearm, and a clear dome helmet carried under her right arm in the full-body panels, captured with true-to-life fabric texture and even soft studio lighting.", "aspect_ratio": "16:9"}
``` ```
Write the file exactly in this compact single-line form: one JSON object, two keys,
no surrounding text, no markdown fence inside the file.

View File

@ -143,14 +143,12 @@ If Subject 1 had a user-provided reference instead, the line would be
## `Image-1.json` ## `Image-1.json`
```json ```json
{"expanded_caption": "A 16:9 horizontal character design board titled \"VOSS\" in a clean technical sans-serif name tag, presenting a female astronaut in her thirties with short black hair and brown eyes, shown in one large full-body view standing in a neutral stance, feet fully visible. She wears a matte-white EVA suit with orange stripe accents along the arms and chest, a clear dome helmet, grey reinforced gloves, and a wrist-mounted scanner with a small amber readout on her left forearm. A role line under the name tag reads \"Deep-Field Survey Specialist\", and a short profile note describes a calm, methodical explorer. Along the bottom edge, a labeled palette strip shows matte white, safety orange, graphite grey, and amber swatches, with thin callout lines pointing to the suit fabric, helmet glass, and scanner housing. Two signature props, the wrist scanner and a compact sample canister, are displayed as small inset items beside the figure on a light grey drafting-paper background with subtle grid lines, so the board reads as a designed character sheet rather than a plain portrait.", "aspect_ratio": "16:9"} {"expanded_caption": "A photorealistic full-color character reference sheet photographed on a clean white background, divided into FOUR panels arranged side by side with equal white spacing between the panels, and no text, no labels, and no borders. Panel 1, the leftmost and narrower panel, is a close-up of the person's face only \u2014 head and shoulders, front-facing, filling most of the panel: a woman in her thirties with short black hair, brown eyes, fair skin, and natural skin texture. Panels 2 through 4 show the same woman full body from head to toe with no cropping: Panel 2 a front view facing directly forward, Panel 3 a side view facing 90 degrees to the right, and Panel 4 a back view facing completely away with the back of her head and body visible. All four panels show the identical person with the same facial features, skin tone, and hairstyle, wearing the exact same clothing: a matte-white EVA suit with orange stripe accents along the arms and chest, grey reinforced gloves, a wrist-mounted scanner with a small amber readout on her left forearm, and a clear dome helmet carried under her right arm in the full-body panels, captured with true-to-life fabric texture and even soft studio lighting.", "aspect_ratio": "16:9"}```
```
## `Image-2.json` ## `Image-2.json`
```json ```json
{"expanded_caption": "A 16:9 horizontal character design board titled \"PIP\" in a rounded technical name tag, presenting a small scout robot shown in one large full-body view in a neutral parked stance, all three wheels fully visible. The robot has a boxy pale-grey shell with soft rounded corners, a single large round blue lens eye centered on its front face, a short antenna on the roof panel, and a triple-wheel base with dark rubber treads. A role line under the name tag reads \"Terraform Scout Unit\", and a short profile note describes a curious, loyal field companion. Along the bottom edge, a labeled palette strip shows pale grey, gunmetal, signal blue, and dark rubber swatches, with thin callout lines pointing to the shell coating, lens glass, antenna joint, and wheel treads. One signature prop, a folded scan projector arm, is displayed as a small inset item beside the figure on a light grey drafting-paper background with subtle grid lines, so the board reads as a designed character sheet rather than a plain portrait.", "aspect_ratio": "16:9"} {"expanded_caption": "A photorealistic full-color character reference sheet photographed on a clean white background, divided into FOUR panels arranged side by side with equal white spacing between the panels, and no text, no labels, and no borders. Panel 1, the leftmost and narrower panel, is a close-up of the machine's front face panel only \u2014 filling most of the panel: a boxy scout robot with a pale-grey shell and a single large round blue lens eye. Panels 2 through 4 show the same machine full body with no cropping: Panel 2 a front view facing directly forward, Panel 3 a side view facing 90 degrees to the right, and Panel 4 a back view facing completely away with the rear panel and short antenna visible. All four panels show the identical machine: a boxy pale-grey shell with soft rounded corners, a short antenna on the roof panel, a triple-wheel base with dark rubber treads, and a folded scan projector arm at the right shoulder, captured with real metal and rubber texture, faint scuffs, and even soft studio lighting.", "aspect_ratio": "16:9"}```
```
--- ---

View File

@ -264,6 +264,21 @@ def main():
if not isinstance(obj["expanded_caption"], str) or \ if not isinstance(obj["expanded_caption"], str) or \
len(obj["expanded_caption"].strip()) < 40: len(obj["expanded_caption"].strip()) < 40:
errors.append(f"{fname} expanded_caption is missing or too short") errors.append(f"{fname} expanded_caption is missing or too short")
continue
cap = obj["expanded_caption"].lower()
if "four panels" not in cap:
errors.append(f"{fname} caption must describe the four-panel "
f"turnaround sheet (face close-up, front, side, back)")
banned = ["cartoon", "anime", "manga", "illustration", "stylized",
"cgi", "3d render", "mascot", "avatar", "name tag",
"swatch", "palette strip", "design board", "drafting paper"]
hits = [w for w in banned if w in cap]
if hits:
errors.append(f"{fname} caption contains illustration/design-board "
f"vocabulary that produces cartoons: {hits}")
if "photorealistic" not in cap and "photograph" not in cap:
errors.append(f"{fname} caption must pin the photorealistic "
f"photographic style for the sheet")
# Lyrics must never appear # Lyrics must never appear
if args.lyrics_file: if args.lyrics_file: