Fix prop confirmation bug, add Wan 2.7 model, refine multiple UI details, improve prop generation quality and aspect ratio, remove text overlays from Asset Center created images, and optimize prop filtering logic

This commit is contained in:
saturn
2026-04-03 22:36:41 +08:00
parent 854b932e67
commit 78b93331b4
136 changed files with 3393 additions and 875 deletions

View File

@@ -3,14 +3,15 @@ You are a key story prop extractor.
Task: identify only key props from the input text for an asset library that must preserve visual consistency across repeated appearances. Be conservative. Return JSON only.
Core definition of a prop:
A prop is a physical object that can exist independently of any specific scene and appears across multiple scenes or timelines. An object qualifies as a prop asset only if a character can "take it away" or "move it to another scene". Most stories have very few props, or even none at all.
A prop is a physical object that can exist independently of any specific scene and appears across multiple scenes or timelines. An object qualifies as a prop asset only if a character can "take it away" or "move it to another scene", and the text provides explicit evidence that the same object is persistently carried, reused, or repeatedly referenced. Most stories have very few props, or even none at all.
Output format:
{
"props": [
{
"name": "prop name",
"summary": "one-line objective prop description"
"summary": "short human-readable prop summary",
"description": "pure visual description for image generation"
}
]
}
@@ -18,11 +19,15 @@ Output format:
Key prop criteria:
1. It must be a real physical object that actually appears in the story.
2. It must be portable — capable of being carried, transferred, or removed from its current scene by a character.
3. It must reappear across multiple scenes or timelines, requiring a consistent visual design.
3. There must be explicit textual evidence that it reappears across multiple scenes or timelines and needs a consistent visual design. Do not infer future recurrence from common sense.
4. It must satisfy at least one of the following:
- characters hold it, use it, fight over it, deliver it, hide it, lose it, or search for it
- it is a key tool, weapon, artifact, piece of evidence, token, key, or clue carrier
- removing it would materially weaken plot comprehension or a key action
- it is a long-term personal item, recurring special equipment, or recurring vehicle tied to a character
5. It must also satisfy at least one uniqueness/continuity condition:
- the object has a non-replaceable identity: ancestral, custom-made, unique, magical, numbered, damaged in a distinctive way, or visually singular
- the text clearly shows the same object reappearing at multiple times or in multiple scenes
- the text clearly shows a character carrying, relying on, or repeatedly searching for the same object over time
Strictly exclude:
1. Ordinary background items, furniture, tableware, food, drinks, daily necessities, and decorations.
@@ -32,6 +37,8 @@ Strictly exclude:
5. Abstract concepts, emotions, powers, roles, places, creatures, and body parts.
6. Scene-fixed facilities — objects that are part of or built into a scene, even if they participate in the plot (e.g. a hacked computer, a smashed window, a fireplace on fire). If the object physically belongs to the scene and cannot be taken away by a character, it is not a prop. These are "scene states" and should be handled by scene descriptions.
7. Scene-standard equipment — if an object is the default fixture of its scene type (a computer in a computer room, a stove in a kitchen, bookshelves in a library, instruments in a lab, screens in a monitoring room), do not extract it.
8. Ordinary replaceable items — even if briefly used by a character, if the story would still work with another generic item of the same kind, it is not a prop. Examples: a fork in a restaurant, a glass on a table, a pen on a desk, a generic phone, a generic umbrella, a generic suitcase, a generic book.
9. One-off action items — if an object is used in only one scene for one action and there is no explicit evidence that it recurs later, do not output it.
Decision bias:
1. A specific-looking noun is not enough; it must have an explicit story function.
@@ -40,27 +47,37 @@ Decision bias:
4. If you are unsure whether it deserves an asset entry, do not output it.
5. Prefer under-extraction. Never output props just to increase the count.
6. Portability test: ask yourself "Can a character put this in their pocket, bag, or car and take it to another scene?" If not, do not output it.
7. Replaceability test: ask yourself "If this were swapped for another ordinary object of the same type, would the story still work?" If yes, do not output it.
8. Continuity test: if the text does not explicitly prove later recurrence or long-term ownership, do not output it.
9. Typical scene items such as a restaurant fork, a wine glass on a table, a lamp in a room, or an office computer should default to not output.
Example judgements (to calibrate your standard):
✅ Extract: a revolver the character carries at all times (cross-scene, portable)
✅ Extract: an evidence envelope (discovered, handed over, appears in multiple scenes)
✅ Extract: a time-manipulating watch worn by the protagonist (core prop, present throughout)
✅ Extract: a black SUV driven by the protagonist (cross-scene transport)
✅ Extract: an ancestral bronze dagger (unique identity, recurring, not replaceable by an ordinary object)
❌ Skip: a computer in a computer room (scene-fixed facility)
❌ Skip: a hacked computer displaying key clues (state change of a scene facility, not portable)
❌ Skip: a surveillance monitor in a monitoring room (scene-fixed facility)
❌ Skip: a refrigerator in a kitchen (scene-standard equipment)
❌ Skip: a rare book in a library (unless the character takes it away and uses it in another scene)
❌ Skip: a fork in a restaurant (ordinary replaceable tableware, even if a character uses it to eat)
❌ Skip: a wine glass on the table (single-scene ordinary object without continuity)
❌ Skip: a standard office laptop (ordinary equipment, not a recurring unique prop)
Output rules:
1. Only output `name` and `summary`.
2. `name` and `summary` must both be non-empty.
1. Only output `name`, `summary`, and `description`.
2. `name`, `summary`, and `description` must all be non-empty.
3. Do not repeat props that already exist in the prop library with the exact same name.
4. Keep names stable and short.
5. Keep summaries objective.
6. Usually output no more than 3 props unless more are clearly all key props.
7. If none exist, return {"props": []}. Returning an empty array is correct in most cases.
8. Replace raw quotation marks inside JSON string values with corner brackets「」.
5. `summary` is for humans only. Keep it short and objective. Do not mention plot function, repeated appearances, character interaction, or camera language.
6. `description` is for image generation only. Keep it purely visual: material, color, shape, structure, quantity relationship, decorative detail. Do not mention usage, plot, action, people, hands, tables, environment, or background.
7. If `summary` or `description` contains phrases like "appears multiple times", "used by the character", "drives the plot", or "in the frame", it is invalid.
8. Usually output no more than 3 props unless more are clearly all key props.
9. If none exist, return {"props": []}. Returning an empty array is correct in most cases.
10. Replace raw quotation marks inside JSON string values with corner brackets「」.
11. It is better to miss borderline candidates than to misclassify ordinary scene items as props.
Input:
{input}