heroyagami/legal-story-image-video

Legal Story Image-to-Video: Abstract Characters × Realistic Backgrounds

Project Positioning

This is an independent image generation and image-to-video workflow. It is based on generating/reusing transparent PNG characters, realistic background images, and localized image-to-video, then assembling the final video through timeline editing. It does not require using any existing code-animation project's engineering, directories, components, scripts, timeline protocols, or rendering gatekeeping.

Usage Entry

Use this when the user provides legal themes, life stories, copy, or reference videos and requests a video featuring "abstract little people placed in a real world." Priority deliverables: character design, character transparent PNGs, scene backgrounds without people, a shot list, image-to-video clips, voiceover, subtitles, and project materials that can be independently edited. The specific tools chosen depend on actual availability.

Visual Rules (Highest Priority)

• Characters and backgrounds must be generated separately and managed separately. Composite them during editing; do not let the generator render characters as real people or in anime style.
• Characters: White round head or extremely simple white face, black dot eyes, simplified eyebrows/mouth, thick black outlines, a few flat color blocks, simple hands and feet. No nose bridge, skin texture, or realistic facial features. When strict reproduction of abstract style is required, no hair or keep only abstract identifiers; character uniqueness relies mainly on clothing, glasses, poses, and accessories.
• Backgrounds: Realistic or highly realistic contemporary Chinese scenes, with real building proportions, furniture, materials, weather, and cinematic lighting. Background images must not contain any people, must not have built-in subtitles, logos, or erroneous legal text.
• Character Assets: Stored as transparent RGBA PNG or transparent SVG; after export, check alpha pixels and edges; do not pretend black/white backgrounds are transparent.
• Emotion: Express using pose, dot eyes/eyebrows/mouth, shot size, lighting, and atmosphere. Warm tones for the hopeful phase, cool tones for the risk phase, restrained for the ending.
• Lawyer and story protagonist must be strictly separated: the lawyer character retains glasses, black robe, and red tie; the homebuyer uses a blue hoodie and backpack; other family roles are numbered separately.

Standard Workflow

1. Topic Selection and Legal Fact Verification: Create research.md, listing public sources, facts and disputes, applicable legal timeframes, and privacy redaction points; mark fictional or composite plots as "situational reenactment."
2. Write the Story: Follow the arc "ordinary life—choice—crisis—consequences—seeking help—legal measures—emotional resolution"; do not mix lawyer narration and first-person protagonist narrative as the same real identity.
3. Character Design: Define each character's character_id, uniform proportions, clothing, and standard expressions based on ../../../references/character-manifest.example.json; establish the front view first, then produce side/back views, expressions, and actions. Do not randomly regenerate for every shot.
4. Transparent Character Image Generation: First produce a single character in a single pose, check the head, clothing colors, line thickness, and transparent alpha, then batch derive variants. One file per character per pose.
5. Realistic Background Image Generation: Generate background plates without characters per shot, preserving the character placement area, default to 16:9 landscape; can also produce 9:16 portrait compositions based on the platform. Record color temperature, light sources, ground plane, and camera position.
6. Shot List and Asset Table: Use ../../../references/shot-list-template.csv. For each shot, specify background_id, character_id, action, duration, shot size, camera movement, sound effects, and whether image-to-video is needed.
7. Static Compositing / Image-to-Video: For normal shots, use the character PNG + static background + gentle push/pull/pan; for shots requiring real motion, use image-to-video or character partial animation for only a few shots. Prioritize controlling consistency of character shape, hands/feet, and clothing color across frames.
8. Voiceover and Subtitles: Generate narration audio from the final script; use subtitle alignment tools to produce calibrated SRT; shot duration follows real audio rather than estimated word count.
9. Editing and Sound: In an independent video editing project, composite characters, realistic backgrounds, camera movements, ambient sound, and restrained BGM. Any editing tool that supports PNG alpha and multiple tracks can be used; do not rely on a specific code-based video framework.
10. Review and Export: Check each shot for character consistency, background realism, lighting and shadow integration, subtitle safe zone, legal information accuracy, and audio integrity; export MP4 and an asset list.

Engineering File Layout (Standalone Video Project)

STORY_PROJECT/
  research.md
  story-script.md
  characters/             # Character transparent PNGs
  backgrounds/            # Realistic scene images without people
  shot-list.csv
  image-to-video/          # Generated video clips requiring motion
  audio/                   # Narration/BGM/ambient sound
  subtitles/               # SRT
  editing/                 # Editing project or EDL
  exports/                 # Final MP4

Base Production Ratio and Camera Movement

• Default: 1920×1080, 16:9, 30fps; change to platform specifications when the user explicitly requests.
• Normal shots can initially be 4–7 seconds, stretch or compress based on narration semantics; gentle push/pull on static backgrounds about 2%–5%, avoid obvious drift.
• Reserve complex image-to-video only for key scenes (walking, turning around, hugging, turning in the rain, etc.).
• Coordinate character and scene perspective, make feet contact the ground and add soft shadows; for rain scenes, synchronize the character's rain exposure state.

Prompting and Acceptance

• Separate prompts for characters and backgrounds are in ../../../references/prompts.md.
• An example character manifest is in ../../../references/character-manifest.example.json.
• This repository provides transparent SVG abstract character master files under assets/characters/vector/, which can be exported as RGBA PNG before editing. The previous 8 PNGs in a standalone ZIP were a conceptual first draft and are not official repository master files. Official production must first pass character design verification.
• For acceptance, first inspect 3 key still frames (hope / crisis / legal help), then produce a 20–30 second test clip, confirm the style, then expand to the full video.

Repository Invocation and Delivery Supplements

• At each start, first read the repository root AGENTS.md and README.md; do not infer tool availability from other repositories.
• Character master files in this repository are located at assets/characters/vector/; they are transparent SVGs; export as PNG as needed for the current editing session before compositing. Generated characters must be compared against the master files.
• First submit a character comparison image, 3 text-free keyframes, and a 20–30 second test clip; only rendered and reviewed videos may be called "finished videos."
• Production instances are uniformly placed outside the repository in independent projects; this repository retains reusable Skills, templates, assets, and necessary version information.

Prevent Degradation into Dynamic PPT (Mandatory)

• Each episode must read ../../../docs/reference-style-qa.md and fill out visual-qa.md item by item.
• Do not treat geometric vector scenes, blurry illustration backgrounds, large title panels, player-style borders, and uniformly standing little characters as "cinematic generated videos."
• Backgrounds must be independently generated / real-space environment materials; characters must remain transparent layers that genuinely interact per shot, including real occlusion, shadows, and different shot sizes.
• FFmpeg is only responsible for final encoding and editing; it does not replace image generation and image-to-video. If no usable image-to-video capability is available, only deliver static images or clearly non-animated samples; do not claim they achieve the original video's animation effect.
• The first version dynamic PPT MP4 produced by A151 on 2026-11-05 has been judged as REJECTED and must not be used as a finished product reference.

Small Characters, Many Shots, Steady Camera (Additional Mandatory Rules)

• When the user says "characters should not take up too much of the frame," they do not mean simply shrinking the character. Characters must maintain real proportions with sofas, hospital beds, tables, chairs, doors, and windows at the same depth; proportions are determined by shot size and true perspective. Any rigid 8%–18% uniform character height limit is forbidden; see ../../../docs/perspective-and-framing.md.
• For the 203-second A151 video, produce 52 independent shots / independent background images; do not reuse three images through cropping/pans to pad 36 shots. Each image must be a real new file with independent composition.
• Default to static camera to prevent shake; moving shots only use slow push/pull within 1.5% and pans within 0.3% in a single direction, with no random oscillation and no frequent reversals. Lock visual center of gravity.
• Related execution materials: ../../../episodes/A151/v3/scene-plan.json, ../../../episodes/A151/v3/image-prompts.md, ../../../episodes/A151/v3/QUALITY_GATES.md.

No Style Drift: Fixed Character Asset Gate

• Reference ../../../references/role-lock-v31.json. Use defined transparent PNG/SVG character assets; do not regenerate characters independently per scene; background images must be absolutely person-free.
• When using ../../../scripts/compose_locked_scene.py for compositing, explicitly provide character identifier, source file, position and perspective reference object, and foreground occlusion mask; this script is only for compositing, not a replacement for AI background generation and image-to-video.
• Every video cover / shot preview must be extracted from an actual produced MP4 frame; do not separately generate a 9-grid collage image and pass it off as a "video screenshot."
• Completion is just that the tool executed successfully; it does not mean the shot has passed visual review. The 23-second test clip of A151 V3.1 has proven that the hybrid workflow can run, but the contact occlusion quality still does not meet the standard.