“Why does this guy get plastic surgery every episode?” — the deadliest comment under any AI comic drama, bar none. Viewers forgive cheap backgrounds, stiff motion, and robotic voiceover; they will never forgive the protagonist growing a new face by shot ten. The human brain’s sensitivity to face consistency crushes its sensitivity to every other visual dimension. Part 2 listed character consistency as a veto item in tool selection; this part is its construction manual — the morphology of failure, how to write a Character Bible, measured Reference Lock data, LoRA accounting, and, having accepted that the gacha exists, budgeting and salvage.
Know the enemy: the four shapes of failure
Consistency failure isn’t mysticism — it has a stable taxonomy with known triggers:
- Cross-shot drift: the protagonist’s features slowly morph between shots. Triggers: segments longer than 5 seconds, frequent cuts, no reference-image anchor.
- Expression drift: a smile becomes a sob within one segment; lip shapes misalign. Triggers: unstable weight on emotion words in the prompt, or two emotion instructions crammed into one segment.
- Wardrobe changes: red dress in the first half, white dress in the second. Triggers: wardrobe not pinned in every prompt, or switching models/style words mid-series.
- Gender/age drift: the model “improvising” on vague descriptions — writing “a girl” is an invitation to reroll the dice.
And a fifth, exclusive to two-person scenes: the chimera effect — hand-holding, hugging, fighting makes the model blend the two characters’ features into each other. The reliable 2026 fix is still generating them separately and compositing in post; don’t expect one prompt to render a contact scene.
The Character Bible: write your character as a spec sheet
All consistency work starts not with a model but with a document. The Character Bible is the project’s constitution, and it deserves 15–20% of the project’s total hours — a repeatedly validated ratio that people who skip it repay doubled in rework.
Four components: the identity sentence (copied into every prompt verbatim, never paraphrased — mechanical is consistent); appearance fields (face shape, eyes, hair, build, signature marks — a mole, a hairpin; these are the cross-shot recognition anchors); the wardrobe lock (pinned per story arc: red dress in episodes 1–8 means red dress in every prompt of episodes 1–8); and the taboo list (every vague word the model improvises on — “girl,” “beautiful,” “young” — banned). Write the Bible before generating the first production image. Do it backwards, and a few hundred wasted images are your tuition.
Reference Lock: the four-step loop and the measured interface table
After the Bible comes the execution loop, four steps: write the Bible → lock references through each model’s interface → generate per-scene (only the scene variable changes) → sanity-check every batch. Step three is the crux: only one line — the scene variable — differs across prompts; identity sentence, style words, wardrobe, and aspect ratio stay frozen.
A May 2026 measurement (same character bible, 24 segments, mean cosine similarity of 8-frame face feature vectors): Veo 3 leads at 0.91 (3 references, $0.50/clip), Hailuo 02 reaches 0.88 (1 main reference, $0.15/clip), Kling 2 reaches 0.86 (up to 4 references, $0.20/clip), Jimeng 3.0 reaches 0.84 (1 reference + identity line, $0.10/clip), Wan 2.7 chained by first frames reaches 0.83. The reference-only approach — no training at all — sits at 85–90% consistency at zero training cost; for anime-style comic dramas, the Vidu/Jimeng/Kling trio covers the main need at a tenth of Veo’s price.
LoRA: worth it or not — the answer scales with your episode count
When the reference approach’s ceiling (85–90%) isn’t enough, LoRA fine-tuning is the next rung: 5–20 multi-angle character images, 15–30 minutes of SDXL training (~¥50 on cloud GPUs, or 6GB VRAM locally), pushing consistency above 95%. But the ledger has to be complete:
- Cost: not just the training run. Dataset prep (insufficient angle coverage means “perfect from the front, broken from the side”), overfitting parameter wrangling, one LoRA per character — a new protagonist restarts the clock. The character-design phase absorbs 15–20% of project hours, and LoRA is the biggest item inside it.
- Return: consistency goes from “good enough” to “boringly stable,” batch QA gets cheaper, and the protagonist becomes an IP asset (merch, spin-offs).
The judgment rule is direct: one protagonist, 30+ episode serialization — worth it; trial phase, anthology format, many supporting roles — not worth it. The middle tier is platform-native character lock (Jimeng/Kling’s paid feature) — trading membership fees for training hours, right for most solo creators under 20 episodes a month. Align the route with Part 2’s budget tiers: ¥0 tier uses free references, ¥500 tier uses platform character lock, ¥2,000 tier adds LoRA.
Reroll budgeting and the salvage ladder
Accept a fact: every 2026 model has a 5–10% failure floor — that’s not your prompting, it’s the current waterline of the technology. The engineered response is pricing failure in and building a salvage ladder:
The cost formula: true cost per kept shot = base price × (1 + failure rate × rerolls). A ¥0.40 base, 8% failure rate, 2 rerolls → about ¥0.46 per kept shot. Every number in Part 2’s per-second table should be multiplied by this coefficient before you budget. The salvage ladder, three rungs: ① reroll the same prompt 2–3 times and pick the best (cheapest, first resort); ② add a 0.3-second crossfade around the broken segment — the audience’s subconscious reads it as a “cinematic transition”; ③ replace the unsalvageable shot with a reaction shot, an empty frame, or a visualizer segment — if the protagonist isn’t on screen, they can’t break. The bottom line: a broken frame the audience never sees costs nothing; a broken frame they see costs the episode.
Action items, then Part 5
- Write a Character Bible: complete all four parts for your protagonist (identity sentence / appearance fields / wardrobe lock / taboo list), then deliberately generate 3 images using one taboo word — watching “improvisation” happen once beats ten tutorials.
- Run the four-step loop: with the same identity sentence + 1 main reference, generate your protagonist in 4 different scenes and check face and wardrobe shot by shot. Record the pass rate — that’s your baseline consistency waterline.
- Test the salvage ladder: take the most broken image from step 2, try both rescues in the edit — the 0.3s crossfade and the reaction-shot swap — and ask three friends whether they can tell. Build your feel for “salvageable rate.”
Part 5 puts your now-consistent character onto the timeline: motion and voice. Image-to-video camera control (first/last frames, camera-move words, motion amplitude), TTS emotion and multi-character voice management, the engineering of audio-visual alignment and lip-sync, and a genuine routing debate — voice-first or picture-first.