YGtemple/video-skills-collection

AI Dynamic Comic Video Production Skill

🏷️ Skill Category Tag

Category: [Comic Drama / Dynamic Comic]

• 🎬 Typical Task Scenarios: Generating campus/youth/healing/crush dynamic comic videos clip by clip based on character design images and storyboard scripts; multi-clip comic drama shorts that require consistent character faces, hairstyles, and clothing across clips; outputting 9:16 portrait or 16:9 landscape mp4 clips to be stitched into a final video.
• 🚫 Non-Applicable Scenarios: Pure product commercials / pure landscape shots / abstract videos without characters (use doubao-creative-video or a landscape skill instead); videos requiring real actors (use jimeng-studio / ai-drama-master-suite realistic branch); long series with dozens of episodes (use long-form drama tools; this skill only produces single-episode short samples).
• 🔗 Upstream / Downstream & Recommended Pairings: Upstream [doubao-creative-drama short-form entry / blockbuster-script-director script / drama-storyboard-director storyboard / script-asset-master asset images] → This Skill (dynamic comic per-shot generation + post-production compositing) → Downstream [jianying-editor fine editing / byted-mediakit-editing transitions & color grading]
• 🧭 Boundaries with Adjacent Skills: Difference from ai-drama-master-suite / jimeng-studio — this skill only handles comic/anime/semi-realistic illustration styles and mandatorily uses image_to_video + reference image driven generation, not realistic live-action; difference from doubao-creative-drama — this skill is the execution/implementation skill on the comic side (takes storyboard and reference images, generates clips per shot, removes watermarks, and composes them), not responsible for project initiation and overall planning.
• 📦 Standard Three-Part Output Responsibility: This skill is responsible for ② General Model Asset Prompts (Comic/Dynamic Comic Version) among the three parts — produces Chinese asset prompts for character four-view sheets, scene background shots, and prop settings according to the General Model Asset Prompt Template, and corresponding English MJ prompts according to the MJ Prompt Template (unified --niji 6 for the comic version); ① Project Outline is produced by the upstream planning entry (doubao-creative-drama / blockbuster-script-director); this skill receives its script and synopsis as input; ③ MJ Prompts are transcribed by this skill from ②, covering key characters/scenes/props assets.
• ⚠️ User Fixed Specifications: Final output has no subtitles, no BGM, and TTS is disabled (dialogue is generated directly by the video model); display the complete prompt before generation; character images must be @ referenced, mandatorily image-referenced, not purely text-generated characters.

⛔ Stage Gate General Rules (Option-Driven · User Confirmation · All Assets Self-Made by User)

This skill by default no longer automatically runs end-to-end complete production — it does not default to "automatically run all clips" nor produce any final art assets for the user. The standard paradigm full text is at /home/user/Doubao/chats/38436370915409154/_work-specifications/stage-gate-general-rules-standard-paradigm.md; the following are the implementation points for this skill (comic / dynamic comic direction), taking priority over any old statements in the text below about "automatic continuous / automatic per-clip full run / AI auto-generates final assets".

Highest Principle (Non-Violable)

1. Prohibited to automatically run a complete production in one go: Must not, after the user gives just a one-sentence idea, automatically complete the entire story → script → assets → storyboard → video → post-production process; also must not, after the user confirms the first clip, automatically run through all remaining clips one by one.
2. At the end of each stage, you must: ⛔ Pause → Give Options (each with a one-sentence highlight + pros/cons + ⭐ recommendation) → Wait for user explicit reply; do not proceed without user confirmation.
3. The user can at any step: reply with a number to select / mix and match multiple options / reply "divergence" for alternative options / directly write revision suggestions.
4. 【All Art Assets Must Be Produced by the User Herself/Himself · Strong Constraint】: All image assets that go into the final video — such as comic character design images (three-view / facial close-ups), scene background images, prop design images, key frame / storyboard reference images — must be generated, selected, and sent back by the user using prompts provided by the AI; the AI must not call any text-to-image (T2I) tool to produce final assets for the user.
   • The AI is still responsible for video generation (image_to_video / I2V): after all assets are sent back and confirmed, the AI generates video by @ referencing user assets, showing the complete video prompt, and proceeding only after user confirmation.
5. No final asset images are automatically output at any stage; if a schematic / reference image is truly needed to aid communication, the AI must first obtain explicit user consent, and the schematic must not be directly used as a final asset in the final video.

Gate Control Unified Output Format (Must be given for each gate point, cannot be omitted)

• ✅ This Stage Completed: 1–3 sentences explaining what was produced in this step.
• Please choose (total N directions): Use a table with columns Number / Option / One-Sentence Highlight / Pros / Cons, and give ⭐ recommended item with reason.
• ⛔ Paused, please reply for me to continue: Clearly state what the user needs to reply (number / confirmation / revision suggestion / "divergence"); only after receiving a clear selection / confirmation will the next stage begin.

Gate Control Rhythm for This Skill (Comic / Dynamic Comic):

Gate	Stage	Options/Actions for User	Condition to Pass
Gate 1	Story Divergence (Before Script)	Give 3–5 mutually differentiated comic story directions (different themes / core settings / protagonist archetypes / main conflicts / opening hook images), each with highlights + pros/cons + ⭐ recommendation	User selects / mixes / diverges and confirms the story before writing the storyboard script
Gate 2	Asset Plan Selection (Plan Only, No Images)	Provide options for character configuration & persona, scene configuration, prop list, key frame direction, etc. (Section A of Asset Self-Production Confirmation Sheet)	User selects / modifies and confirms the asset plan
Gate 3	All Assets Self-Made by User (Key, Batch-by-Batch Gate Confirmation)	AI produces two sets of prompt packs (General Chinese Model version + MJ English Comic version --niji 6) + 3D parameters (--stylize 200–600 / --weird 0 / --chaos 0–10 + fixed --seed) + asset list and production order (character anchor → scene bg → props → key frames) + production suggestions; AI does not call T2I tools to output images; user generates, sends back, and confirms one by one	User generates, sends back, and confirms characters/scenes/props/key frames one by one before @ referencing them in video prompts
Gate 4	Batch Video Generation	AI each time first shows the complete video prompt (@ referencing user-confirmed assets), explains the reference image list and parameters; user confirms before calling image_to_video; done in batches, not defaulting to automatically run all clips	User confirms prompt batch by batch before calling video model; after generation, shows and user confirms "continue" before next batch
Gate 5	Post-Production / Delivery	Options for watermarked removal pure crop, xfade transitions, audio mixing, platform transcoding, cover (cover image also self-made by user)	User confirms item by item

Do not enter video generation until all characters/scenes/props/key frames are sent back and confirmed; video and post-production are confirmed batch by batch / item by item, not accumulated for automatic run.

Hook and Causality (Two Hard Creation Requirements, Integrated into the Six-Point Quality Gate of Storyboard Script Below)

• ① Hook Combined with Image (Visible, Not Relying on 'Telling'): The opening hook must be a visible concrete image / action / scene, relying on the first frame / first few seconds of visual impact or visual suspense to grab the viewer, not relying on narration / dialogue / subtitles / titles to 'tell'; the hook image directly relates to the core conflict / suspense. In story divergence and storyboard, "opening hook" must be filled with specific hook image (who is doing what / what abnormal image appears), not abstract concepts.
• ② Causality (A→B→C, Eliminate Rundown): The plot advances via causal chain, not a list of events connected by "and then… and then…"; the anomaly / suspense thrown by the hook must have a cause; each character action has motivation and consequences; foreshadowing is the "cause", payoff is the "effect". When self-checking, question each storyboard shot "what caused it and what does it lead to?", delete paragraphs that have no causal relation to the main story.

Three-Part Output in Batches According to Gate Control Rhythm (Do Not Give All at Once)

• ① Project Outline / Storyboard Script: Produced / finalized only after "story confirmed".
• ② General Model Asset Prompts, ③ MJ Comic Prompts: Provided in batches according to user generation progress — give for character when character is confirmed, give for scene when scene is confirmed; do not give all in advance at once.

Relationship with Existing Specifications (Overlay, No Conflict, No Deletion)

Retain and overlay: 🏷️ Real / Comic classification tags; 📋 Storyboard script six-point self-check; 📦 Standard three-part; 🎛️ Asset 3D (--stylize / --weird / --chaos + fixed seed); ⚠️ No subtitles, no BGM, TTS disabled, display complete prompt before generation, character/asset images mandatory @ and image reference. Two points changed by this general rule: ① Stage progression changed from "automatic continuous / automatic per-clip full run" to "pause per gate, give options, wait for confirmation"; ② Art assets changed from "AI generates images" to "all final assets made by user, AI only gives prompts / 3D / list / suggestions, does not call T2I". All other capabilities remain; video generation (image_to_video) is still executed by the AI.

I. Applicable Scenarios

• Generating AI dynamic comic/short drama videos clip by clip based on existing character design images (three-view/close-up/multi-scene references) and storyboard scripts
• Themes such as campus youth, healing, crush, daily life with character dialogue and emotional interaction
• Multi-clip video production requiring strict consistency of character faces, hairstyles, clothing, and scenes
• Output as 9:16 portrait or 16:9 landscape mp4 video clips, which can later be stitched into a complete short film

Not Applicable: Pure product commercials, pure landscape shots, abstract concept videos without characters, videos requiring live actors.

---

II. Core Principles (Non-Violable)

1. Reference Image Driven: Must use image_to_video, passing character design images + scene reference images; prohibit text_to_video from generating characters from scratch
2. Per-Clip Confirmation: Show each generated clip to the user immediately; only proceed to the next after user confirmation; prohibit batch generation
3. Show Prompt Before Generation: Before each call to the video generation tool, must first display the complete prompt text to the user; only after user confirmation call the tool to generate; prohibit generation without displaying prompt
4. No BGM and TTS Disabled: Sound design includes only character dialogue + ambient sounds + sound effects; absolutely prohibit any background music/piano music/BGM; dialogue/narration must be voiced directly by the video generation model, disable all TTS speech synthesis dubbing — if dialogue is mispronounced/unclear, revise the prompt and regenerate, do not use TTS as a fix
5. No Subtitles: Prohibit subtitles, titles, explanatory text, or any other overlaid text in the image
6. Unique Characters: The number of characters in each shot must match the storyboard; prohibit generating a third character or character duplication (e.g., "two female leads")
7. Unique Limbs: Each character has only two hands and five fingers; prohibit extra hands, extra fingers, or twisted arms (e.g., "three hands")
8. Continuity Priority: The scene, lighting, character state, and prop positions of adjacent clips must strictly carry over from one to the next
9. Voice Consistency: Voice characteristics of male/female lead must be unified across the entire film; voice characteristics must be annotated for each line of dialogue
10. Crop Margin: When generating, must consider the later watermarked removal crop amount; compose with safety margin; key content should not be flush with the edge

---

III. Character Asset Management

⛔ P0 Must Read: Before entering any asset phase, first completely Read the SKILL.md of script-asset-master (live: /home/user/.doubao/agent_mode/workspace/.user_skills/script-asset-master/SKILL.md; stable copy: /home/user/Doubao/chats/38436370915409154/optimized-skills/script-asset-master/SKILL.md); cannot skip, cannot execute from memory. Fixed order of asset phases: Read script-asset-master (P0) → Output asset plan options → User confirms plan → AI gives two sets of prompts + 3D parameters + asset list/order + production suggestions → User self-makes and sends back → Confirm item by item before @ referencing.

📦 Three-Part · Part ② (Comic Asset Prompts) Production Responsibility: This skill produces Chinese asset prompts for character four-view sheets, scene background shots, and prop settings according to the General Model Asset Prompt Template (/home/user/Doubao/chats/38436370915409154/general-model-asset-prompt-template.md), with the style paragraph uniformly switched to "dynamic comic/illustration/semi-realistic anime/cel shading"; simultaneously produces corresponding English MJ prompts according to the MJ Prompt Template (/home/user/Doubao/chats/38436370915409154/MJ-prompt-template.md), with the comic version uniformly adding --niji 6. Character images provided by the user must be @ referenced, one image per character without duplicate assignment; prohibit purely text-based generation of characters.

🎛️ Three Control Dimensions (Must be marked for each asset, none can be missing): Next to the main text of each asset prompt, must list three values/ranges and briefly explain the reasoning; cannot only give the prompt main text.

• ① Stylization (--stylize / Realistic ↔ Stylized Range): MJ uses --stylize (0–1000); recommended for comics under --niji 6: --stylize 200–600 (medium~high), which gives more illustration/dynamic comic artistic feel; too high will deviate from reference and weaken character consistency.
• ② Weirdness (--weird / Creative Oddity Range): MJ uses --weird (0–3000); recommended for comics: --weird 0 (default) to maintain character credibility and consistency; only start from 250–500 for experimental exploration.
• ③ Diversity (--chaos + --seed / Seed Variation): MJ uses --chaos (0–100) + fixed --seed; recommended for comics: --chaos 0–10 with fixed seed to ensure reproducibility; can briefly raise chaos to 10–25 when exploring compositions, must revert for final outputs.
• ⚠️ High weird/chaos significantly sacrifices character consistency and reproducibility — for high-consistency assets like character four-view sheets and cross-shot keyframes, uniformly use low weird, low chaos, fixed seed; can temporarily loosen for exploratory outputs, must revert to recommended ranges for final outputs.
• Each asset must have an annotation line next to it: ▶ Control Dimensions: --stylize [value] (Reason: )｜--weird [value] (Reason: )｜--chaos [value] + --seed [value] (Reason: )

⛔ Asset User Self-Making Boundary (Confirmation Point): In this phase, this skill only produces the above two sets of prompts / 3D parameters / asset list and production order / production suggestions; must never call T2I tools to produce final images of characters/scenes/props/key frames for the user; all final art images are generated, selected, and sent back by the user using the prompts; after item-by-item confirmation by the user, this skill will @ reference them in video prompts (one type per confirmation, no duplicate assignment). If a schematic image is truly needed to aid communication, must first obtain explicit user consent, and it must not be directly used in the final video.

3.1 Types of Reference Images Required

Type	Purpose	Minimum Quantity
Character Full-Body Three-View	Uniform style, full-body proportions, overall character design	1 per person (can be on the same image for both genders)
Character Facial Close-Up	Facial features, hairstyle, expression baseline	1 per person
Scene Reference Image	Spatial layout, light direction, furnishings, atmosphere	1 per core scene
Character Scene Interaction Image	Character's pose, lighting, and proportion in that scene	As needed

3.2 Reference Image Numbering Management

All reference images are numbered sequentially by input order as @image1, @image2, @image3..., and must correspond one-to-one with the order in image_reference_url_list in the prompt.

Suggested Reference Image Asset Table:

| Number | Image Content | Type | Purpose |
| Image1 | Male and female lead full-body three-view | Character design | Uniform style, full-body proportions |
| Image2 | Male and female lead dual close-up | Character face | Facial features, hairstyle |
| Image3 | Female lead solo close-up | Character face | Female lead facial features, hairstyle |
| Image4 | Male lead in library scene | Scene + Character | Library space, male lead's pose |
| Image5 | Library full view | Pure scene | Spatial layout, lighting, furnishings |

---

IV. Reference Image Citation Standard (Mandatory)

4.1 Citation Format

At the beginning of the prompt, must list the specific reference content for each image one by one, in the format:

Reference @image1's [specific reference content], reference @image2's [specific reference content], reference @image3's [specific reference content].

4.2 Character Image Citation Must Specify

• Face: Facial features, eyes and brows, nose shape, face shape, skin tone
• Hairstyle: Hair color, length, curly/straight, bangs, hair volume
• Clothing: Style, color blocking, school badge position, innerwear, pants, shoes
• Body type: Height, build, proportions
• Example: Reference @image2's male lead facial features (cool and handsome, clear brow bone, black short hair texture) and female lead facial features (gentle and delicate, big eyes, black long curly hair + air bangs)

4.3 Scene Image Citation Must Specify

• Space: Bookshelf arrangement, table/chair positions, window position, overall layout
• Lighting: Light source direction, color temperature, intensity, shadow direction
• Furnishings: Props, decorations, floor material
• Atmosphere: Overall color tone, emotional feel
• Example: Reference @image5's library scene spatial layout (wooden bookshelf arrangement, large window on the left, reading area by the window) and afternoon warm golden light direction

4.4 Number of References

• Each clip must include at least 2 reference images (1 character + 1 scene, or 2 character images)
• Clips with characters together must include a reference image containing both characters (e.g., dual close-up)
• When the scene changes, must pass in the corresponding scene reference image
• Maximum 10 images (tool limit)

---

V. Prompt Structure Template (Mandatory)

The prompt for each clip must strictly follow the following order, with none missing:

[Paragraph 1: Reference Image Citation]
Reference @image1's [specific use], reference @image2's [specific use], reference @image3's [specific use].
(List one by one; for character images, specify face/hairstyle/clothing; for scene images, specify space/lighting/furnishings)

[Paragraph 2: Continuity Anchor]
Carry over from previous clip: [scene state at the end of the previous clip, character position/pose/expression, prop position, light direction].
(For the first clip, write "Establishing opening scene"; for subsequent clips, must write the carry-over content)

[Paragraph 3: Image Content]
[Detailed description: Scene environment + character action + expression change + key props + details]
[Must specify how many characters are in the image and who they are; prohibit extra characters]

[Paragraph 4: Camera Movement]
[Starting shot]→[Movement method]→[Ending shot], clearly state pan/tilt/zoom/track/fixed.

[Paragraph 5: Lighting and Style]
[Light source direction, color temperature, depth of field, art style, quality requirements]

[Paragraph 5.5: Composition Safety Zone]
[Clearly state that key content (character faces, hands, props) is strictly within the central 55% of the image; distance from top of head to top edge is at least 22% of height; distance from sole/chin to bottom edge is at least 18%; at least 250px safety margin on left and right sides; not flush with the edge, not in the corners, not cut by the image edge. Reserve sufficient blank space around the image edges to preserve space for later large-crop watermarked removal.]

[Paragraph 6: Dialogue]
Character name (Voice characteristic): "Dialogue content"
(Each line of dialogue must be annotated with character name and voice characteristic; male lead low, magnetic, slow; female lead gentle, clear)

[Paragraph 7: Sound Design]
[Only ambient sounds + sound effects; prohibit writing any BGM/background music/piano music/score]

[Paragraph 8: Prohibitions]
No subtitles, no on-screen text, no BGM background music; only [number of characters] characters in the image ([character names]); prohibit a third character or character duplication.

---

VI. Continuity Specification

Between adjacent clips, the following elements must remain consistent to avoid errors:

6.1 Scene Space

• Bookshelf arrangement, table/chair positions, window position, door position unchanged
• Prop positions (e.g., which table the book is on, whether the chair is pushed in) unchanged
• Spatial direction (left/right/front/back) not flipped

6.2 Lighting

• Light source direction consistent (e.g., afternoon sunlight entering from the left window throughout)
• Color temperature consistent (warm golden, no sudden change)
• Shadow direction consistent
• The entire film takes place in the same time period; no jump from daytime to dusk

6.3 Character State

• The pose/expression/position at the end of the previous clip is reasonably carried over into the next clip
• Clothing details consistent (zipper position, school badge side, sleeves rolled up or not)
• Hairstyle state consistent (messy from wind, tied up or not)

6.4 Props

• Key props (e.g., The Little Prince) must be consistent in whose hands or where they are placed
• The appearance, color, and position of props unchanged

6.5 Continuity Anchor Writing Example

Carry over from previous clip: Female lead Shen Zhiyi just found *The Little Prince* in front of the bookshelf in the literature section and reached for it; the afternoon warm golden sunlight shines in from the left window; the female lead stands in front of the bookshelf looking up; male lead Xu Zhouzhou walks over from the other side of the bookshelf.

---

VII. Character Voice Standards

Character voices must be consistent throughout the entire film; each line of dialogue must be annotated with voice characteristics:

Character	Voice Characteristic	Speed	Tone	Volume
Male Lead	Low, magnetic male voice	Somewhat slow	Calm, restrained	Moderate to low
Female Lead	Gentle, clear female voice	Natural	Gentle, with youthful shyness	Moderate
Female Lead Inner Monologue	Soft, breathy voice	Somewhat slow	Inner thoughts, distinct from dialogue	Soft

Writing Format

Male lead Xu Zhouzhou (Low, magnetic male voice, somewhat slow, calm and restrained tone): "You want to read this too?"
Female lead Shen Zhiyi (Gentle, clear female voice, gentle tone with a hint of shyness): "Mm... I didn't finish it last time."

Prohibitions

• Do not use AI robotic voice
• Do not confuse male and female voices
• Do not have voice characteristics change suddenly within the same clip
• Do not use narration to replace character dialogue (unless the storyboard clearly specifies an inner monologue)

7.1 Dialogue Flow Standards (Crucial!)

Common Issues

• Break: Lack of logical connection between two lines of dialogue, character conversation jumps abruptly
• Unclear Expression: Meaning of dialogue is vague; the audience cannot understand what the character is saying
• Rhythm Break: Dialogue is not back-and-forth; characters talk to themselves
• Abrupt Information: New information suddenly introduced (e.g., essay competition) without setup

Dialogue Design Principles

1. Logical Connection: Each line of dialogue must respond to the previous one, or naturally lead to the next; cannot be disjointed
2. Back and Forth: Dialogue is interaction, not monologue; what A says B responds to; what B says A picks up
3. Information Setup: When introducing new information (e.g., essay competition, grade ranking), there must be prior setup or natural transition
4. Fit the Situation: Dialogue must fit the current scene and character emotions; cannot shout in a quiet library
5. Fit the Personality: Male lead speaks little and is calm; female lead is gentle and shy; dialogue style must match
6. Conversational: Use natural conversational language; not formal written language; no stiff self-introductions

Break Fix Example

❌ Break Version:

Female lead: "It is you! I'm Shen Zhiyi, from Class 1, Grade 11. I always hear your name at school assemblies."
Male lead: "Shen Zhiyi... I know. Your essay 'The Old Bookstore' from the last writing competition was really good."

Issue: After the female lead says "school assemblies," the male lead suddenly jumps to "writing competition" — lack of connection, break.

✅ Flowing Version:

Female lead (Surprised): "It is you! I'm Shen Zhiyi, from Class 1, Grade 11. No wonder you looked familiar — at every school assembly I see you go up for an award."
Male lead (Pauses slightly, serious): "Shen Zhiyi... I remember you. That essay 'The Old Bookstore' you wrote in the last city-wide writing competition — I really liked it."

Fixes:

• Female lead adds "No wonder you looked familiar" to connect "It is you" and "school assemblies"
• "see you go up for an award" is more specific and natural than "hear your name"
• Male lead's "I remember you" is more humane than "I know"
• "city-wide writing competition" is more specific than "last writing competition"
• "I really liked it" has more emotion than "was really good" and fits the secret crush subplot

Dialogue Check List

Before generation, must check line by line:

• Does each line of dialogue respond to the previous one?
• Is new information set up?
• Is the dialogue back-and-forth?
• Does the dialogue fit the character's personality?
• Is the dialogue conversational and not stiff?
• Is there any break or jump?
• Is the line length appropriate for the clip's duration?

Speed and Rhythm Standards (Crucial! Based on Measured Experience)

Common Issues:

• Speech is too fast, lines crowd together and are hard to hear
• Multiple character voices overlap, next line starts before previous finishes
• Deliberately slowing down speech causing unnaturalness
• Adding too long pauses between lines causing rhythm break

Correct Practice:

1. Normal Speed: Reference the speed annotations of already confirmed normal clips (e.g., clips 3, 4, 5); do not deliberately slow down or speed up
2. Male Lead Slightly Slow: Fits the cool, calm, restrained personality; annotate as "slightly slow speed," not "extremely slow"
3. Female Lead Normal Speed: Annotate as "gentle, clear female voice," no need to additionally write slow
4. Dialogue Naturally Alternating, Not Overlapping: Write "dialogue naturally alternates, no overlap," do not write "pause for 2-3 seconds between each line"
5. Do Not Deliberately Add Pauses: Do not write (pause 3 seconds) between lines; let the AI handle dialogue rhythm naturally
6. Concise Lines: Each line should not be too long; 4 lines of dialogue suggested at 12-14 seconds to give ample room for generation
7. Duration Matching: More lines → increase duration (e.g., 14 seconds); fewer lines → use 10-12 seconds; do not pack too many lines into a short time

❌ Incorrect Writing:

Dialogue (extremely slow speed, pause 2-3 seconds between each line):
Female lead (extremely slow): "..."
(Pause 3 seconds)
Male lead (extremely slow): "..."

✅ Correct Writing:

Dialogue (naturally alternates, no overlap, normal speed):
Female lead Shen Zhiyi (Gentle, clear female voice, gentle and shy tone): "..."
Male lead Xu Zhouzhou (Low, magnetic male voice, slightly slow speed, calm and restrained tone): "..."

---

VIII. Anti-Drift / Anti-Multi-Character Constraints (Crucial)

8.1 Common Issues

AI video generation is prone to:

• Character duplication (e.g., two female leads in one shot)
• Facial drift (facial feature deformation, face swap)
• Clothing changes (color, style suddenly change)
• Extra characters (random passerby appearing)
• Prop deformation (book turning into something else)

8.2 Mandatory Constraint Wording

In the Prohibitions section of each prompt, must write:

There are only [X] characters in the image, namely [list of character names]; prohibit a third character, prohibit character duplication, prohibit extra people.
Character facial features, hairstyle, and clothing must match the reference images; prohibit deformation, face swap, or changing clothing color and style.

8.3 Image Content Section Must Also Specify Number of Characters

When describing the image, specify the number of characters at the beginning:

There are only two characters in the image: Male lead Xu Zhouzhou and female lead Shen Zhiyi. The male lead stands on the other side of the bookshelf...

8.4 Special Attention for High-Risk Scenes

The following scenes are especially prone to multi-character drift and require extra constraints:

• Reaching for something (hand close-up easily generates extra hands or people)
• Close interaction between two people in frame
• Back/side profile (easily generates an extra front-facing character)
• Mirror/glass reflection (reflection easily generates extra character)

Handling Method: In the prompt for these scenes, the Prohibitions section repeatedly emphasizes the character count constraint, and in the Image Content section, confirm the number of characters sentence by sentence.

8.5 Anti-Multi-Hand / Anti-Limb Drift (Crucial!)

Common Issues

Hands and limbs are most prone to errors in AI video generation:

• Multi-Hand: A character with three hands (e.g., two hands on bookshelf + one hand holding book)
• Multi-Finger: Abnormal number of fingers (six fingers, fused fingers)
• Abnormal Arm: Extra arm generated, arm twisted, arm passing through body
• Hand Deformation: Hand turning into a claw, fingers melting, hand proportion distortion
• Prop Clipping: Hand passing through book, arm passing through bookshelf

Mandatory Constraint Wording

In the Prohibitions section of each prompt, must write:

Each character has only two hands and five fingers; prohibit a third hand, multi-finger, twisted arm, hand deformation, or limb clipping. Hand movements are natural and reasonable; finger count is normal.

Image Content Section Must Also Specify Hand Movements

When describing hand movements, must clearly state what each hand is doing:

The female lead holds three books against her chest with her right hand, and gently runs her left fingertips over the spines of books on the shelf. (Clearly state the division of labor for two hands; prohibit a third hand)

Special Attention for High-Risk Scenes

The following scenes are most prone to multi-hand/hand deformation and require extra constraints:

• Reaching for something/reaching for a book (hand close-up)
• Holding a book while touching the bookshelf (when the two hands' roles are not clearly differentiated)
• Passing/receiving something (two hands close together)
• Fingers tracing a page/book spine (finger close-up)
• Hand in pocket while holding something (hand position switching)
• Multiple hands in the same frame

Handling Method:

1. In the Image Content section, clearly state the position and action of each hand sentence by sentence
2. In the Prohibitions section, repeatedly emphasize "each character has only two hands"
3. If the action is complex, consider splitting into two shots to avoid too many hand actions in a single shot
4. For actions like "holding a book + touching the bookshelf," clearly write "one hand holds the book, the other hand touches the bookshelf"; do not use vague description

---

IX. Prohibitions List

The Prohibitions section of each prompt must contain all of the following:

1. No Subtitles — No subtitles, titles, or explanatory text in the image
2. No BGM — No background music, piano music, score, or melody
3. No Watermark Text — No watermark text like "AI Generated" (to be cropped in post-production)
4. Fixed Number of Characters — Only the roles specified in the storyboard; prohibit a third character or character duplication
5. Fixed Number of Limbs — Each character has only two hands and five fingers; prohibit a third hand, multi-finger, twisted arm, or hand deformation
6. Character Consistency — Face, hairstyle, and clothing must match the reference images; prohibit deformation or face swap
7. Scene Consistency — Spatial layout, light direction, and furnishings must not contradict the reference images and previous clip
8. No Extra Props — Do not include props not mentioned in the storyboard

---

X. Model Selection

Model	Permission	Duration	Resolution	Recommended Scenario
seedance_2.0_fast	Free	5-15 sec	1280×720 / 720×1280	Default use, regular users
seedance_2.0	Member only	5-15 sec	Higher	High quality required and have membership
seedance_2.5	Member only	5-30 sec	Higher	Need clips >15 seconds

Selection Rules

1. Regular users directly use seedance_2.0_fast, do not try 2.0/2.5 (will be rejected)
2. Keep single clip duration within 5-15 seconds, recommend 8-12 seconds
3. Generate at most 2 video clips at a time, recommend generating one by one
4. Aspect ratio based on needs: 9:16 (portrait short drama), 16:9 (landscape cinematic)

---

XI. Generation Process

11.1 Preparation Phase

⛔ Prerequisite for entering this phase (Confirmation Point): The following character design images / scene reference images are assets self-made by the user using the above prompts, sent back and confirmed item by item; this skill does not generate these images for the user. Do not enter 11.2 per-shot generation until all assets have been sent back and confirmed.

1. Confirm character design images are complete (three-view + facial close-up)
2. Confirm scene reference images are complete (at least 1 per core scene)
3. Confirm storyboard script is complete (each shot has shot size/image/dialogue/duration/camera movement/sound)
4. Storyboard Script Six-Point Self-Check (Mandatory Quality Gate, pass before entering per-shot generation) — see below '11.1.1 Storyboard Script Self-Check List'
5. Establish reference image asset table (number + content + use)

11.1.1 Storyboard Script Self-Check List (Comic/Dynamic Comic Version · Quality Gate)

Source: Short Drama Script Creation Self-Check List (/home/user/Doubao/chats/38436370915409154/short-drama-script-creation-self-check-list.md). For comics, wording is slightly adjusted to "visualization/storyboarding"; six core key points are retained. After the storyboard script is completed and before per-shot video generation, must check each item; if any item fails, revise the script/storyboard before generating.

• ① Conflict Design (Visualized): The core conflict makes the audience feel involved and is visible through images/actions/props (not explained by narration); each conflict point has specific foreshadowing (early shots set it up, later shots pay it off); at least 2 conflicts; conflicts simultaneously serve the three purposes of shaping the protagonist, driving the plot, and elevating the theme
• ② Character Profile (Establish Character via Shots): Write characters based on situation, do not apply labels; describe personality traits and the problems they face; establish characters in shots through mannerisms, actions, and habitual details (a signature action/expression throughout the entire film)
• ③ Scene and Transition (Shot Connection): Design specific transitions to avoid hard cuts (transition via light/shadow/prop/state change); each shot conveys the time-space context; for the same scene at different times, use changes in lighting, prop positions, and character states to show the passage of time
• ④ Narrative Structure (Multi-Line Interweaving + Causal Chain): The story contains multiple lines — main plot + family line + social relationship line, each line intersects and drives others in key shots, rather than being stacked in parallel; the plot advances via causal chain A→B→C, not an 'and then…and then…' rundown; each shot should be able to answer "what caused it and what does it lead to"; delete paragraphs with no causal relation to the main plot; character actions have motivation and consequences; foreshadowing is the cause, payoff is the effect; twists are unexpected yet reasonable in context; the anomaly / suspense thrown by the hook must have a cause (can be revealed later)
• ⑤ Opening Hook and Interaction (Visualized · Retention): The opening hook must be a visible concrete image / action / scene, relying on the first frame / first few seconds of visual impact or visual suspense to grab the viewer, not relying on narration / dialogue / subtitles / titles to 'tell', and directly relates to the core conflict / suspense; not using a bauble unrelated to the main content; use a hot button / focus / high-conflict image to capture the audience at the opening; set discussion points (controversy / suspense / resonance) throughout the film to trigger comment interaction
• ⑥ Audience Perspective Self-Check (Nitpick Per Shot): Switch to an ordinary audience perspective and review the storyboard table again; answer shot by shot "what can be optimized? Does this shot make me want to continue watching?"

One-Sentence Process: First establish characters and scenes → arrange ≥2 conflicts and foreshadow them in the storyboard → use the main plot to string together emotion/social relationship lines → design transitions and the opening hook shot → embed discussion points → finally, from the audience perspective, nitpick shot by shot.

11.2 Single Clip Generation Process (⛔ Display Complete Prompt Batch by Batch, Generate After Confirmation, Do Not Default to Automatically Run All Clips)

1. Write the prompt for that clip according to the storyboard (strictly follow the template in Chapter V, including reference image citations, continuity anchors, composition safety zone, anti-multi-hand/anti-multi-person constraints)
2. Display the complete prompt text to the user, explain the reference image list and parameters, and wait for user confirmation
3. After the user confirms the prompt, select the corresponding reference images, pass them in order into image_reference_url_list
4. Call image_to_video (default model seedance_2.0_fast); recommend generating one clip at a time, or at most two clips in parallel (must be two consecutive clips verified to be independent)
5. After generation, display the video to the user; after user confirms "continue" or "satisfied," proceed to the next clip
6. If the generated clip is unsatisfactory, analyze the issue (character inconsistency / limb deformation / scene error / expression not matching expectation / extra person / etc.), modify the prompt and regenerate; do not accept a failed generation

---

XII. Concrete Prompt Examples

Example 1: Female Lead Solo Clip (Showing Volume of The Little Prince)

Reference @image1's male and female lead full-body three-view (campus uniform style, school badge on left chest, Shen Zhiyi's long black curly hair + air bangs, Xu Zhouzhou's black short hair, school uniform style), reference @image3's female lead solo close-up (gentle and delicate facial features, big eyes, long black curly hair + air bangs), reference @image5's library full view (wooden bookshelf arrangement, large window on the left, reading area by the window, afternoon warm golden light).

Carry over from previous clip: Female lead Shen Zhiyi just found *The Little Prince* in front of the bookshelf in the literature section and reached for it, afternoon warm golden sunlight shines in from the left window, the female lead stands in front of the bookshelf looking up.

Image: Only female lead Shen Zhiyi stands in front of the bookshelf in the literature section. The camera is positioned slightly behind and to the right of her shoulder. She holds a copy of *The Little Prince* in both hands, the book open, her gaze gently looking at the book cover, with a hint of a smile at the corners of her mouth. The afternoon sunlight shines through the window, casting on her side profile and the book pages, creating a warm glow. The female lead's long curly hair falls over her shoulders, her expression is soft and focused. The book title *The Little Prince* is clearly visible. The female lead stands among the bookshelves, the surrounding books stacked neatly. (Only one person: Shen Zhiyi)

Camera movement: Shoulder-level shot from behind → Slow push-in to front-facing close-up → Fixed on female lead's facial expression. Starting at medium shot, slowly pushing in to close-up (waist-up to bust-up), focal point on the female lead's face and the book cover.

Lighting and style: Warm golden side light from the left, shallow depth of field (book in focus, background bookshelf slightly blurry), semi-realistic anime illustration style, 4K quality, detailed.

Composition safety zone: The female lead's face is strictly within the central 55% of the image; the book is in front of the chest, within the central area; the top of Shen Zhiyi's head is at least 22% of the image height below the top edge; the bottom of the picture starts from her chest area, not cut at chin or neck; at least 250px safety margins are reserved on the left and right sides; key content is not flush with the image edge.

Dialogue: None (pure action and expression display).

Sound design: Only the soft rustle of pages turning of *The Little Prince* in the quiet library; subtle ambient sound of distant footsteps.

Prohibitions: No subtitles, no on-screen text, no BGM background music; only one character (Shen Zhiyi) in the image; no third person or character duplication; the female lead has only two hands holding the book; prohibit a third hand or extra fingers; hand movements are natural and reasonable; finger count is normal.

Example 2: Dual Character Meeting Clip

Reference @image1's male and female lead full-body three-view (campus uniform style, school badge on left chest, body proportions), reference @image2's male and female lead dual close-up (facial features, hairstyle), reference @image5's library full view (spatial layout, lighting, furnishings).

Carry over from previous clip: Female lead Shen Zhiyi is looking through *The Little Prince* in the literature section, standing in front of the bookshelf, afternoon sunlight shining in from the left window, the book open in both hands.

Image: Only two people: Male lead Xu Zhouzhou and female lead Shen Zhiyi. The male lead walks from the other side of the bookshelf, slightly surprised to see the female lead in front of the bookshelf. The female lead hears footsteps, looks up from the book, meeting the male lead's gaze. Their eyes meet across the bookshelf. The afternoon sunlight passes through the gap between the bookshelves, casting on both their faces. The male lead stands by the bookshelf, the female lead holds the book, the two facing each other. (Only two people: Xu Zhouzhou and Shen Zhiyi; no third person or passersby)

Camera movement: Start with camera behind the male lead's shoulder, seeing the female lead in the distance from across the bookshelf → Slowly pan to medium two-shot → Slight push-in to waist-level shot. Starting at full shot, slowly push-in to medium two-shot, finally locked on both characters' facial close-ups.

Lighting and style: Warm golden side light from the left, light emanating from the bookshelf gap, shallow depth of field (characters in focus, background slightly soft), artistic and warm overall, semi-realistic illustration style.

Composition safety zone: Both characters' faces are strictly within the central 55% of the image; the distance between them is appropriate in the frame, not too tight or too loose; at least 250px safety margins on left and right; the top of the male lead's head is at least 22% of image height from the top edge; key content is not flush with the edge.

Dialogue:
Male lead Xu Zhouzhou (Low, magnetic male voice, somewhat slow, calm and restrained tone): "You like fairy tales too?"
Female lead Shen Zhiyi (Gentle, clear female voice, gentle tone with a hint of shyness): "Mm... I guess you could say so."

Sound design: Only the ambient sound of the quiet library; the soft summer breeze outside the window.

Prohibitions: No subtitles, no on-screen text, no BGM background music; only two characters (Xu Zhouzhou and Shen Zhiyi) in the image; prohibit a third character or character duplication; each character has only two hands; prohibit multi-finger, twisted arm, or hand deformation; hands and limbs are natural and reasonable; images not allowed to have extra arms or hands.

---

XIII. Post-Production Processing

13.1 Watermark Removal (Crop)

• Method: Use clip_video_editing tool, select "crop" (cropping)
• Crop Range: The final 16:9 videos need to remove built-in watermarks at the bottom; the specific crop range is determined according to the generated video
• Timing: The generated video clips need to go through watermarked removal before they can be spliced
• Safety Margin: When generating, the safety margin must be sufficient (especially at the top and bottom); the center 16:9 area should not contain content that needs to be preserved

13.2 Transition Effects

• Default: Use clip_video_editing tool, select "xfade" (cross dissolve) for transition between clips
• Duration: 0.5 seconds
• When: After crop watermarked removal, before splicing

13.3 Audio Mixing

• Use clip_video_editing tool, select "audio_mix" (audio mixing)
• Merge all consecutive video clips into one
• Ensure audio does not have splice/cut issues

13.4 Export Specifications

Parameter	9:16 Portrait	16:9 Landscape
Resolution	720×1280	1920×1080
Frame Rate	30fps	30fps
Codec	H.264	H.264
Bitrate	8-12 Mbps	12-16 Mbps

---

XIV. Parameter Configuration Table

Parameter	Default Value	Adjustment Range	Notes
clip_duration	10 seconds	5-15 seconds	Recommend 10-12 seconds for dialogue scenes
resolution	1280×720 (16:9)	720×1280 / 1280×720	Choose based on platform
image_reference_url_list	[]	2-10	Must include at least 1 character + 1 scene
batch_size	1	1-2	Recommend generating one at a time
frame_rate	30fps	24-30fps	Recommend 30fps for smooth motion

---

XV. Common Issue Handling

Issue	Cause	Solution
Character face deformation	Reference image does not match prompt description	Re-describe facial features; increase image reference weight
Extra character appears	Prompt does not strictly constrain character count	Add "only [number] characters" in Prohibitions
Clothing color changes	Prompt description is imprecise	Add specific color and style description; reference image clothing details
Multi-hand/multi-finger	Prompt lacks limb constraint	Add "each person only has two hands" constraint
Scene lighting inconsistent	Light direction not specified	Add light direction and color temperature description
Prop missing or changes	Not described in detail	Describe prop shape, color, and position
Video length too short	Storyboard is too tight	Extend clip_duration or adjust storyboard content
Dialogue is not clear	TTS sound quality, or character dialogue area mismatch	Increase dialogue scene character proportion; reduce background noise; ensure character mouth visible

---

XVI. Quality Standards

16.1 Output Quality

• Each clip: minimum 5 seconds, maximum 15 seconds
• No subtitle watermark or manufacturer watermark in the image (cropped)
• Smooth natural motion without frame skipping or stuttering
• Consistent characters: face, hairstyle, clothing match the reference images
• Consistent scene: spatial layout, lighting, furnishings consistent
• Clear dialogue, matching the character and scene
• No extra character or background person appears
• Hands are normal (two hands, five fingers counts, no hand deformation)
• Props are consistent in shape and position

16.2 Story Quality

• Clear core conflict is set up and resolved
• Characters have clear personalities and growth (visible through actions/expressions)
• Plot has clear causal chain, no 'and then' run-on
• Opening hook visually captures viewer attention
• Emotional layers: culminating emotional peak, not flat throughout
• The ending has aftertaste (resonance / suspense / open ending)

---

XVII. Tool List

Function	Tool	Notes
Video generation	image_to_video	Core tool, mandatory use
Video editing	clip_video_editing	Crop watermarked removal, transition, audio mixing
Image generation	(User self-service)	User uses prompts to generate via Midjourney/other tools; this skill does not call

---

XVIII. Quick Start Checklist

Refer to the Stage Gate General Rules for the full step-by-step list; the following is the simplified core operation flow (must follow gate-by-gate confirmation rhythm):

• 1. Read the full SKILL.md of script-asset-master first
• 2. Gate 1: Give story directions → user confirms the story → only then write storyboard
• 3. Write storyboard script → pass six-point self-check
• 4. Gate 2: Output asset plan options → user confirms the plan
• 5. Gate 3 (User self-makes assets): Provide two sets of prompts + 3D parameters + list/order + suggestions → user generates, sends back, confirms one by one
• 6. Gate 4 (Generation): Each clip: display complete prompt → user confirms → generate → user confirms → next clip
• 7. Gate 5 (Post-production): Watermark removal → transitions → audio mixing → export → user confirms delivery