FAQ
I. Generation Failure / Abnormal Status
Q: The video shows “Generation Succeeded”, but when clicked it cannot play / won’t open?
A: The generation actually failed but the status was incorrectly shown as success (✅ fixed). Just regenerate. If it still fails, click “Problem Feedback” and send the link.
Q: The video keeps spinning (over half an hour) and never finishes generating?*
A: Service fluctuation or API timeout. After recovery, the task will be queued automatically; wait a while. If nothing happens for a long time, send the problem link.
Q: “Network connection failed” error?
A: Transient network fluctuation. Just regenerate once.
Q: Images won’t generate (GPT Image 2 / Channel 1/2/3)?
A: ① Triggered safety interception → reword the prompt or switch channels; ② API call failure → switch to Channel 3 or re-enter the page.
II. Sensitive Words / Copyright Interception
Q: The prompt keeps showing “Sensitive interception”, but I checked and it involves no politics/porn/drugs?
A: ① Involves political figures (e.g., “Harris”) → rename; ② Content too bloody → soften the description; ③ System false positive → retry on another channel; if confirmed as a false positive, you can request an exemption.
Q: Audio keeps flagging “sensitive” — domestic works but overseas doesn’t?
A: Overseas audio review is stricter. ① Strengthen the prompt with “no BGM”; ② mute the reference audio before uploading; ③ retry 1-2 times.
Q: Generation fails with “output video may be related to copyright restrictions”?
A: The character resembles a celebrity or well-known IP. Modify the asset’s appearance and add differentiating features (accessories / special hair color) to avoid it.
Q: The same asset used to work, but now suddenly flags copyright/sensitive?
A: It was recognized as a look-alike or IP infringement. Modify the character asset and regenerate.
Q: I don’t know which exact sentence in the prompt is sensitive. How do I find out?
A: Click “AI Assistant Diagnosis” to locate the sensitive word, If no sensitive words are detected, copy the issue link and ask customer support to contact the official team.
III. Character / Persona Issues
Q: The character in the generated video doesn’t match the design sheet / characters are swapped / female lead and supporting actress mixed up?
A: ① Image order scrambled (✅ fixed); using existing assets is unaffected; ② add the character name to the prompt and reference with @character; ③ delete redundant/incorrect reference images.
Q: The character was not generated per the design (clothing / hairstyle / features don’t match)?
A: Add the design description to the prompt or use @character to reference the design image; you can also extract a frame, edit the image, and re-import it.
Q: One character generates several different-looking people?
A: Multiple design sheets were switched for the same character. Best practice: one design sheet per character; confirm before switching.
Q: Pupil color is wrong / style changed?
A: Submit the problem link for official investigation of parameters and model output.
IV. Dialogue / Dubbing / Voice Issues
Q: I provided English dialogue, but the output speaks Chinese?
A: The duration is too short, so the dialogue can’t be finished and falls back to Chinese. ① Increase video duration; ② emphasize “speak in English” in the prompt; ③ confirm the reference voice is set correctly.
Q: The audio reference is female, but the output is always male?
A: The reference voice setting is wrong. Adjust the reference audio source bound to the character and retry.
Q: No dialogue was added, but the character speaks anyway (in Chinese)?
A: Occasional model glitch. Check the reference audio, regenerate, or declare “no dialogue / silent”.
Q: There’s a strange repeating tail sound / two people saying the same line a few tenths of a second apart?
A: Dubbing glitch. Submit the problem link; if necessary, switch to a domestic model to draw the audio.
Q: Dialogue is read wrong (“why” pronounced as the Chinese equivalent, etc.)?
A: Emphasize “speak in English / read the original text” in the prompt, wrap dialogue in {}, and forbid the model from rewriting.
V. Scene / Visual Issues
Q: The scene prompt says night, but it always generates daytime?
A: Strengthen the night constraint — emphasize the time twice at both the beginning and end of the prompt; it’s an occasional issue, just retry.
Q: Frequent scene changes / inconsistent background within the same segment?
A: Keep multiple angles in the background reference image, fix the spatial description, and avoid switching camera angles within the same shot.
Q: The playing-card asset always generates as “face-up cards” instead of card backs?
A: Upgrade the negative constraint to a strong visual instruction: only show uniform card backs, never render any card face.
Q: The design sheet has no seat belt, but the generation always has one?
A: Extract a frame, erase the seat belt with a P-edit, and import it as the reference image to regenerate.
Q: The people in the scene image are too small / the scene is too large. How to scale?
A: Adjust the “spherical radius”, “camera field-of-view angle”, or the character scale value. A true panorama is required.
VI. Assets & Material Management
Q: Two assets with the same name appear in the library (e.g., two “Chelsea”)?
A: The frontend now prevents duplicate names. Historical data needs technical investigation; if it affects usage, rename it.
Q: Image / character / scene review is very slow, stuck on “Under Review”?
A: There was a review-service stall previously.
Q: Scene images submitted directly for scene review show “contains real person”?
A: The scene was missed or misjudged in review. Submit it as “Other” type to pass review first, or wait for official re-review.
Q: Favorited assets don’t have image-generation prompts?
A: The official plan is to bring out prompts when editing. Prompts are original attributes and cannot be modified by default.
Q: Can the single reference-image upload limit be raised to 40MB?
A: The SD limit is 30MB. Compress 4K large images to under 30MB (change to 16:9 2K), or wait for the official compression feature.
Q: I accidentally deleted an asset (clicked the trash can). Can I recover it?
A: It cannot be recovered. Please confirm before deleting.
Q: One show has multiple branches, and other groups’ assets leaked into our project shell?
A: Projects should be isolated. If wrong data was uploaded, delete it directly; or duplicate the project to split it, and use favorites to pin items.
VII. Feature Operation & Experience
Q: Can we add a “copy material to multiple episodes” feature?
A: Put it in the project asset library and it can be reused across episodes. The official version already supports custom categories and favorites.
Q: After using the replace feature, characters are scrambled in the generated video (Image 1/2/3 mappings mixed up)?
A: Local file concurrent upload order got scrambled (✅ fixed). Replacing with existing assets is unaffected.
Q: After closing the video window, sound continues / a blank frame remains?
A: The window now hides instead of closing after becoming modeless (✅ fixed). It will be changed back to a modal window or stop on close.
Q: Remove-subtitles is easy to trigger accidentally — it starts erasing with one click?
A: It has been moved into the playback dialog to avoid accidental clicks.
Q: Can I import a panorama as background in the Director’s Stage and scale it?
A: Adjust the spherical radius / field-of-view angle to change scene size; adjust the scale value if characters are too small. A panorama is required.
Q: Can frame extraction output white-model / line art?
A: Frame-extraction white-model conversion has been added for precise positioning control; AI recompute gives higher quality.
Q: Can the global unified settings (default parameters) be customized? Changing them every time is tedious?
A: Custom global unified settings have been added — you can save common templates.
Q: Can the @ + number image-selection feature in storyboards be improved?
A: Already optimized: @ inserts into the storyboard, numbers prioritize shortcut keys, panel UI reworked, and @character auto-brings audio.
VIII. Script Parsing
Q: Script parsing failed / an episode is greyed out and unclickable / can’t be parsed?
A: Script format issues (e.g., too many blank lines). Edit and fix, then re-trigger parsing; fixing the anomaly auto-triggers re-parsing.
Q: Can a project that hasn’t started production re-upload the script for parsing?
A: Yes, re-uploading the script for re-parsing is now supported for projects that haven’t started production.
Q: Where do I upload the script / create a new drama shell for production?
A: Upload the script when clicking “Start Production”; creating a new drama shell also goes through the start-production flow.
IX. Model Selection & Channel Strategy
Q: Which model is best for image generation? What’s the difference between Channel 1/2/3?
A: There’s no “best” — choose by scenario:
- GPT Image 2 (Channel 1): best overall quality; suited for character design sheets and key scenes; but pricier, prone to noise in complex scenes, and strictest review;
- Nano Banana 2 (Channel 2): best value; suited for batch generation with clean output;
- Jimeng / Doubao etc. (Channel 3): looser review; suited as a fallback for edgy/dark scenes, but slightly lower quality. Suggestion: Channel 1 for design sheets/key scenes, Channel 2 for batch/routine, switch to Channel 3 as fallback when Channel 1 blocks you.
Q: Domestic model or overseas SD 2.0 for video generation?
A: Domestic takes about 5 minutes; overseas is 20-30% slower. Use domestic for quick validation, overseas for specific styles. Note overseas audio review is stricter and queues are longer.
Q: How to choose between Seedance 2.0 Mini and Fast?
A: Mini: 2× speed, ~62% cost, same quality — suited for high-frequency mass production; Fast: suited for quick validation.
Q: How is the HappyHorse 1.1 model?
A: Newly integrated model; suits specific styles. Switch per project needs.
X. Production Workflow & Standards
Q: How should project assets be categorized to be standard?
A: Store by “Character” / “Scene” / “Prop” categories. The official version supports category stats and batch moving.
Q: Multiple episodes in one project share the same scene/character — how to avoid duplicate uploads?
A: Put it in the project asset library and all episodes share it. Custom categories and favorites are supported.
Q: What’s the difference between old mode and new mode? What if I’m still using old mode?
A: Old mode has no new features and is being phased out. We recommend switching to the new production mode.
Q: Is the lip-sync feature still available?
A: It has been migrated under “Material Management”. Please download and back up old-mode videos in advance.
Q: The final-video review UI is awkward and the progress bar doesn’t match?
A: The layout has been optimized. If issues remain, send a screenshot as feedback.
Q: After starting production, can I re-upload the script for parsing?
A: If production hasn’t started, you can re-upload; if it has started, manually trigger re-parsing after editing the script.
XI. Collaboration & Permissions
Q: How do I assign a project to the production team?
A: Assign it when clicking “Start Production”. Author display and episode labels have been added.
Q: What’s the difference between preset roles and this show’s roles?
A: Preset roles = cross-project shared templates (team collaboration module); this show’s roles = concrete characters in the current work (3.3 Character Design). Manage them separately.
Q: Someone else uploaded materials into my project. How do I clean it up?
A: Projects should be isolated. If data got mixed in, delete it directly, or duplicate the project to split it.
XII. System & Network
System freeze / failed to fetch storyboard info / page won’t load?
A: Service fluctuation or internal bandwidth saturated. Wait and retry, or Ctrl+F5 hard refresh.
Error opening storyboard production (e.g., “Yr.has is not a function”)?
A: Chrome extension conflict. Disable the offending extension at chrome://extensions/, or switch browsers.
The video preview stutters at the 3rd second?
A: Network fluctuation — the 3rd second loads a second chunk. It’s normal once the network stabilizes.
XIII. Prompt Engineering (Industry Experience)
Q: How detailed should a prompt be? Does writing too long make it worse?
A: Use the three-part structure: ① Tone-setting section (style/ratio/constraints) ② Content section (scene → action → camera movement → dialogue, segmented along the timeline) ③ Constraint section (repeat core constraints). 800–1500 characters is the sweet spot.
Q: Chinese or English in prompts?
A: Use Chinese for descriptions (looser review, richer semantics) and English for technical instructions (better adherence). Mixed writing is recommended.
Q: How do I write negative prompts (don’t want XX) effectively?
A: Negative constraints have low adherence. Rewrite them as strong positive visual instructions. E.g., ❌ “no seat belt” → ✅ “plain clothing without accessories”.
Q: The same character looks inconsistent across different storyboards?
A: ① Fix the reference images (front + three-quarter side) and @character-reference them in every storyboard; ② unify the style description at the start of each prompt; ③ avoid extreme angles.
Q: How do I control transitions between storyboards? The model always hard-cuts.
A: ① Keep the same background reference image within the same scene; ② state the new scene at the start of the prompt for different scenes; ③ avoid writing “camera pans from A to B” inside a single storyboard.
XIV. Storyboard & Storyboard Design (Industry Experience)
Q: How long should a storyboard be?
A: 5–10 seconds is optimal. Under 3s characters tend to deform; over 15s backgrounds tend to drift. Test the line’s reading duration before setting the storyboard length.
Q: How many storyboards for the same plot segment?
A: Simple dialogue 1–2, emotional conflict 2–4, action scenes 3–6. Don’t cut too finely — each extra storyboard adds another chance of character breakage.
Q: How to use reverse shots (shot/reverse-shot)?
A: Use the Director’s Stage / white model to fix positions, generate the front shot + reverse shot reference images (consistent background, mirror-symmetric positions), and use them as the storyboard background reference.
Q: How exactly do I use the Director’s Stage (3D scene control)?
A: ① Import a panorama as the background; ② place character positions; ③ set camera position; ④ export as reference. Fine-tune same-scene storyboards based on it — greatly improves consistency.
XV. Art Style & Style Consistency (Industry Experience)
Q: The art style is inconsistent across different storyboards?
A: ① Write the style description in the global unified settings; ② reference the same style anchor image in every storyboard; ③ don’t mix models within one project.
Q: What style suits AI short dramas? Which is easiest to get good results?
A: Easiest: hyper-realistic CG and warm-toned idol-drama style. Hardest to control: pure realistic live-action and 2D anime. New teams should start with hyper-realistic CG.
Q: I mixed different image models in one project and styles don’t match?
A: Avoid it if possible. If mixing is unavoidable: ① unify style with image-to-image; ② unify color grading in post.
XVI. Cost & Efficiency (Industry Experience)
Q: How much does an AI short drama cost? How to control the budget?
A: ① Use Channel 2/3 for image drafts first, Channel 1 for the final design; ② use domestic/Fast for video validation first, final version after locking; ③ reuse audio whenever possible.
Q: How long does a 1-minute AI short drama take to produce?
A: A skilled team takes about 2.5–4.5 days per minute. The bottleneck is review stalls and character-consistency rework; once the pipeline is smooth, efficiency improves by ~50%.
Q: How do I judge that a storyboard is “good enough” and stop drawing?
A: ① Character is 80% like the design sheet; ② key costume/scene features are correct; ③ the line is read correctly. Pass if these hold. If it fails over 5 times, the prompt has a problem.
XVII. Asset Production Pipeline (Industry Experience)
Q: How should character design sheets be prepared for more stable videos?
A: A design sheet is a “good reference image”, not a “pretty image”. ① 3 images per character (front / half-body / full-body); ② clean background, neutral expression; ③ clear differences between characters.
Q: Should scene images be real photos or AI-generated?
A: AI-generated scene images work better — the model references them more naturally. Real photos tend to inherit noise and texture; convert to CG style first if using them.
Q: When should I use the frame-extraction trio (paint-over, tagging, annotation, line-art/white-model)?
A: ① Paint-over to remove unwanted elements; ② tagging to prevent character mix-ups; ③ white-model to control positioning (recommended flow: white model first → confirm position → image-to-image for the final look).
Q: Can reference videos (real filmed footage) be uploaded directly as reference?
A: Yes. ① Must pass review; ② the model inherits the camera movement and lighting style; ③ recommended duration 2–10 seconds.
XVIII. Review & Compliance (Industry Experience)
Q: Why do some words that look fine keep getting intercepted by AI?
A: Common false positives: political figure names, IP character names, gory/violent words, edgy descriptions. Solution: rephrase with vague Chinese descriptions or switch to Channel 3.
Q: What’s different between overseas and domestic review standards?
A: Overseas is more sensitive to copyright/BGM; domestic is stricter on politics/porn but more lenient on gore. Use domestic for copyright-risk content, overseas for pure marketing.
Q: How do I avoid “look-alike” celebrities?
A: ① Don’t design sheets “resembling XX celebrity”; ② differentiate character features (freckles / special hair color / glasses, etc.); ③ if repeatedly flagged, switch channels and regenerate the design sheet.
Appendix: High-Frequency Universal Tips (Official Best Practices)
- No-BGM audio template: emphasize twice at both the beginning and end of the prompt: “no background music, no score, no BGM throughout — only vocal dialogue + ambient sound effects”; wrap sound effects in
<>and dialogue in{}, never use**（）**to write music descriptions (that notation is detected as BGM); the API can setgenerate_audio=falseto disable from the source. - Overseas vs domestic models: overseas SD 2.0 is slower but better in some visual aspects; domestic average generation is optimized to about 5 minutes. If image channels (GPT Image 2 / Nano Banana 2 / Jimeng, etc.) fail, switch channels and retry.
- Seedance 2.0 Mini vs Fast: Mini is 2× speed, ~62% cost, improved motion quality, same image quality — suited for e-commerce/marketing/UGC high-frequency production; Fast is suited for quickly validating creative drafts.
- Problem feedback: for any anomaly, prefer clicking “Problem Feedback” to copy the link and send it to the group — locating issues by link is faster than screenshots (the link is sometimes in the clipboard; paste it into a text editor and re-copy to avoid it being overwritten by an image format).