Failure → Fix Gallery

The 14 most common ways AI-video prompts go wrong — and the exact fix for each. Every case is grounded in faq.md (full Q&A) and methodology.md (the 5-stage structure).

Each card: ❌ Symptom / 🔍 Cause / ✅ Fix — with a concrete before→after prompt snippet wherever one helps.

中文版 →

How to use this page

Skim the symptoms, jump to the one that matches your output, apply the fix. For the why behind each fix, follow the link at the bottom of the card to the relevant FAQ entry or methodology stage.

Sections:

1. Faces & IP
2. Realism & CG-look
3. Motion & timing
4. Audio
5. Endings

1. Faces & IP

Uploaded face photo gets blocked by moderation

❌ Symptom: Your reference photo is rejected or flagged before generation even starts.

🔍 Cause: Face moderation on AI-video platforms keeps tightening. Direct full-face uploads have the highest rejection rate.

✅ Fix: In order of escalation —

1. Try several different photos.
2. Feed the photo to Doubao with the commercial-portrait describer template to generate a "photorealistic colored sketch" version, then upload that instead.
3. Design the character to avoid full-face exposure entirely: helmet, mask, robot, back-of-head shots, mannequins with big sunglasses. (This is the real reason Zombie Scavenger's lead is a robot.)

→ faq.md — Faces / IP / blocked content

Prompt gets flagged for IP / copyright

❌ Symptom: A prompt naming Kamen Rider, Gundam, Pacific Rim, Iron Man, or Kai'Sa gets blocked by Seedance 2.0.

🔍 Cause: Seedance's IP-rights filtering is aggressive and blocks real film / character names.

✅ Fix: Strip the names; keep the design language as description.

- Iron Man red-and-gold armor, repulsor blast
+ atom-punk retro-futurist red-and-gold suit, palm energy burst
- Black Sun aesthetic, gritty tokusatsu
+ gritty dark battle-damaged aesthetic, gritty tokusatsu

Also works: delete a few characters or punctuation marks without changing meaning, or swap in synonyms for the same look.

→ faq.md — Faces / IP / blocked content · prompts/README.md

Face / identity drifts between shots

❌ Symptom: The character's face shape or expression changes from clip to clip; the same person doesn't read across cuts.

🔍 Cause: Weak Stage 2 lock-down. If you don't explicitly pin facial features, the AI drifts.

✅ Fix: Lock the face hard in Stage 2 (Character & scene):

Face: Reference uploaded photo. Features, face shape, hairstyle 100%
      preserved. No beautification. Maintain [wound / scar / blemish].

For transformations, go even harder — keep bone structure identical across the change:

Preserve the same bone structure, jaw geometry, and facial features
100% identical before and after transformation.

Give each key visual element (face, expression, key prop) its own dedicated reference photo so the AI doesn't blend them.

→ methodology.md — Stage 2

2. Realism & CG-look

Output looks like cheap plastic CG, not live-action

❌ Symptom: Faces and surfaces read as glossy, fake, game-render CG instead of film.

🔍 Cause: Two usual culprits — (1) over-beautified / filtered reference photos that already look fake, (2) low-quality reference images with no real photoreal or 3D-render texture.

✅ Fix: Use clear, large, un-filtered headshots — ideally bare-face (no movie puts beauty filters on its actors). Only feed reference images when the source is genuinely high-detail photoreal or a 3D render. For dynamic shots (transformations, combat), drop references entirely and let the AI free-form from pure text.

→ faq.md — Picture quality / texture / realism

Reference image leaks its art style into the output

❌ Symptom: Your output picks up the style of the reference (anime, drawn, CG render) instead of just its design.

🔍 Cause: AI binds a reference image's texture and style information together — poor reference quality leaks unwanted aesthetics into the result.

✅ Fix: Don't attach a reference unless it is genuinely photoreal or a high-quality 3D render. For motion shots, omit references entirely — Mx-Shell's Kamen Rider armor came entirely from text, which also made every output uniquely different. Reserve references for character/scene continuity only.

→ methodology.md — Principle #1 · faq.md — References vs. pure text

Video feels stiff and game-CG — no "presence"

❌ Symptom: The shot is technically fine but feels artificial and locked-off, like game cinematics rather than a handheld film moment.

🔍 Cause: Missing the "breathing" handheld float, and no explicit realism anchors.

✅ Fix: Add realism anchors in Stage 1 (cinematic | hyperreal | no game-CG feel), and in Stage 4 always include the breathing line:

Breathing: Handheld shot. Throughout, maintain an extremely subtle,
           breath-like camera float to enhance presence.

Name a real camera + lens so the AI has a concrete visual anchor — e.g. IMAX film camera + Panavision C-series (35mm, f/4) for epic shots, or Sony Venice + Canon K-35 for gritty sci-fi.

→ methodology.md — Stage 3 & 4

A specific camera move won't execute

❌ Symptom: Wrong focal length, wrong angle, or the orbit goes the wrong direction.

🔍 Cause: Vague descriptions like "cinematic" or "dramatic" carry no precision. The AI needs explicit camera models, focal lengths, and movement vocabulary.

✅ Fix: Be precise in Stage 4 — camera + lens, shot size, angle, motion type, and rhythm:

- Cinematic dramatic camera move around the hero
+ IMAX + Panavision C 35mm. Medium close-up, low-angle 30° from left.
+ Orbit clockwise, very slow, with a 0.1s reactive micro-shake.

Vocabulary to pull from — shot sizes (wide / medium / close-up), angles (low / high / level / over-shoulder), motions (push-in / orbit / follow / FPV), rhythms (very slow / steady / sudden / reactive).

→ methodology.md — Stage 4 · faq.md — Camera & composition

Color grade falls apart in post (banding, noise)

❌ Symptom: Pushing color correction even a little causes banding and noise.

🔍 Cause: AI-generated video has a low color bitrate — far less grading headroom than real-camera footage.

✅ Fix: Lock your color tone upstream, at the image/prompt stage (Stage 3, Color & tone), and keep that section consistent across every shot. Make the big aesthetic choice in the prompt (Hollywood teal-orange, grey-blue, retro warm-orange + sea-salt blue…). Do only minor transitions in CapCut — don't try to fix color in post; fix it in the prompt.

→ faq.md — Duration / long videos / editing

3. Motion & timing

One-take video collapses past ~15 seconds

❌ Symptom: A transformation or single-shot runs long and quality breaks down mid-sequence.

🔍 Cause: Seedance struggles with one-shot generations much longer than 15s. The sweet spot is 5–15s by scene type.

✅ Fix: Keep transformations ≤ 15s per shot. For longer sequences (>15s combat), generate the first 15s, then use Xiaoyunque web client's "Continue this video" to extend. This segmented approach beats one 30s+ shot. Mx-Shell built Kai'Sa as a 15s version plus a separate 5s extension.

→ faq.md — Duration / long videos / editing · prompts/kaisa-transformation.md

Prompt exceeds the character limit; won't submit

❌ Symptom: A long, detailed multi-shot prompt hits the character ceiling and can't be sent to Seedance 2.0.

🔍 Cause: Detailed segmentation and multi-shot scripts run past the mobile client's limit.

✅ Fix: Switch from the mobile app to the Xiaoyunque web client — it has a higher character ceiling. Still too long? Trim ruthlessly: drop redundant descriptors, merge related clauses, abbreviate Scene lines. (Mx-Shell's Pacific Rim prompt explicitly notes it needs the web client.)

→ faq.md — Word limits / rerolls · prompts/pacific-rim-gundam.md

Two clips don't connect — feels like two videos

❌ Symptom: Cut two clips together and it reads as two separate shots, not one continuous action.

🔍 Cause: No motion-direction continuity. The audience can't read one action flowing into the next.

✅ Fix: Connect by motion-direction continuity, not by matching static frames. Zombie Scavenger example: the robot throws a bomb that blasts the horde and pushes him out one side of frame; the next shot, he enters from that same side. Design shots with deliberate exit/enter edges to build logical spatial flow.

→ faq.md — Duration / long videos / editing

A great shot generated once, then won't reproduce

❌ Symptom: A combat or weapon-charge shot lands beautifully once, then every reroll fails to reproduce the camera work.

🔍 Cause: AI generation is non-deterministic. Mx-Shell's Flame Demon camera move was an accident — the AI didn't actually follow the instruction, and he can't reproduce it either.

✅ Fix: Accept randomness as part of the process. Treat one prompt as a ticket to draw, not a deterministic recipe — invest in reroll selection, not prompt polish. Mx-Shell's typical counts: high-difficulty shots 20+ tries, simple shots 2–3. Zombie Scavenger burned ~400 images

• 200+ video shots to land ~40 final clips. **Curate the output; don't

chase prompt perfection.**

→ faq.md — Word limits / rerolls / non-reproducibility

4. Audio

AI adds background music you didn't ask for

❌ Symptom: The model scores your clip with non-diegetic music when you only wanted production audio.

🔍 Cause: Default AI-video models add music unless told otherwise.

✅ Fix: Always specify it in Stage 3 (Atmosphere & look):

Sound: No score. Production audio only.

Mx-Shell used this formula across every prompt (Zombie Scavenger, Kamen Rider, Kai'Sa, Metal Gear charge-combat, Pacific Rim, Cyber-Wuxia).

→ methodology.md — Stage 3

5. Endings

The ending is an FX pile-up and reads cheap

❌ Symptom: Stacked explosions, light bursts, and particle effects at the end overwhelm the frame and read as amateur.

🔍 Cause: Excess FX at the climax overstages the moment. Mx-Shell deliberately avoids it.

✅ Fix: Practice restraint — end quietly. No dialogue, no explosion, no blinding light. Just the character in the moment: standing, breathing, in smoke, with one distant element (a meteor crossing the sky). A restrained ending reads as more elevated than an FX-heavy one.

"No dialogue. No explosion. No blinding light. Just a figure standing in the smoke, a meteor crossing the distant sky." — Mx-Shell, Methodology Principle #3

→ methodology.md — Principle #3

See also

• faq.md — the full Q&A this gallery distills.
• methodology.md — the 5-stage structure every fix refers to.
• prompts/ — Mx-Shell's complete original prompts.