Skip to main content

茶米兔的AI世界

茶米兔的AI世界
FANCY RABBIT · AI, TESTED BY HAND
中文EN

WORKFLOW

How Night Huntress Kept the Same Face for 8 Episodes: The Face-Reference Workflow

By Fancy Rabbit · Published 10 月 15, 2026

How Night Huntress Kept the Same Face for 8 Episodes: The Face-Reference Workflow

A locked face reference, a real production workflow, and the limits that remain.

AI series rarely die from bad plots. They die when the lead walks into episode three wearing a different face. This post breaks down the exact workflow we used on Night Huntress season one: how the reference image is made, how the prompts are written, and how the same lock works for video — so your character can survive a whole season too.

ENVIRONMENT
Night Huntress season-one pipeline, 8 episodes (16 shots × 10 seconds per episode, 1280×720)
TOOLS
Image generation + first/last-frame image-to-video; reference file refs/yeer-face-closeup.jpg
TESTED ON
Workflow locked 2026-10-02; face-reference A/B test 2026-10-14
COST
See the cost table below

Face drift is a workflow problem, not luck. Crop one face close-up from the locked character art, reference it in every single generation — and the lead keeps the same face. Ours did, for all 8 episodes.

Why do AI characters keep changing faces mid-series?

Across the first three episodes of Night Huntress, face drift was our most expensive source of redos: the same lead, Li Ye’er, would have sharp eyeliner in one shot and a softer, sweeter stranger’s face in the next. When we locked the production workflow on 2026-10-02, “character consistency” was written in as priority number one — a ranking bought with three episodes of rework.

The mechanism is simple once you see it. Describing a character in words means re-introducing her to the model on every shot. No matter how detailed the description, the model’s reading of it shifts slightly each time — and across dozens of shots, small shifts compound into a different person. Our conclusion: stop describing the face. Show it one.

What is a face reference, and how do you crop one?

A face reference is a close-up of the character’s face, cropped from the locked character art and saved as one single file that every generation in the season references. For Night Huntress, we cropped it from Li Ye’er’s original locked portrait (refs/yeer-face-closeup.jpg). From then on, every generation featuring her — stills, video, covers — cited the same image, with the prompt stating that her features follow the reference.

Three rules govern the crop:

  1. It must come from the locked art. Never build a reference by chaining generations — cropping a generated frame to seed the next round compounds the error, generation after generation, until the face has quietly become someone else.
  2. The face must be large and complete. Both eyes, brows and lips clearly in frame. This image is the only “real person” the model gets to know.
  3. Shrink the file before uploading. We resize references below 500KB (large character-bible images below 200KB) — oversized reference files trigger upload errors, a wall we hit early and wrote into the workflow document.
Face close-up reference of Li Ye'er, lead character of Night Huntress
The one face reference shared by the whole season. One small trap: oversized reference files fail to upload — resize below 500KB first.

How do you write the prompt so the face actually locks?

Every Night Huntress prompt carries two fixed components. First, a verbatim style opener — “thick painterly anime illustration with visible brushstrokes, NOT photorealistic, NOT 3d render, NOT live-action” — which nails the art direction down and explicitly excludes photorealistic, 3D and live-action looks. Second, the feature anchor: the prompt states that the lead’s features match the reference image, with no room for “similar enough”.

Anything else the model likes to improvise gets written down per shot, in plain words. Rain is our standard example: only fully outdoor shots may show falling rain, while interior shots state “NO falling rain anywhere” and allow wet reflections and puddle shine only. It is the same logic as the face lock — whatever you don’t pin down, the model will happily invent.

Does the lock hold for video, not just stills?

For video we add one more input: a style reference image (style-ref-shenye-original.jpg) to keep the brushwork consistent in motion. On 2026-10-14, during full-length pipeline testing, we ran a controlled comparison on a two-character noodle-stall shot (test-shot02): same script, same first frame, one variable — whether the face reference was supplied.

Version Setup What happened to the face
v1 First + last frame drive, no face reference Face drifted softer and sweeter; proportions moved
v2 First frame + face reference as the second input, plus a [face lock] instruction Sharp eyeliner and deep-red lips back to the locked design; proportions matched; four sampled frames, same person
v3 v2 plus a style lock (brushwork and face shading locked too) Features and makeup stable throughout; face shading still slightly smoother than the reference — not fully solved

The result went straight into the workflow document: any shot with a face toward camera now defaults to “first frame + face reference” mode. The lock holds for video — but only if you pick the right input mode, rather than copying the stills recipe unchanged.

A/B test frames: v1 without a face reference on the left, v2 with the face reference on the right
Frames from the A/B test (second 2), same shot, same first frame. Left, v1 without a face reference — the face drifts soft and sweet. Right, v2 with the reference — back to the locked design (the red tassel on v2’s earring is a small deviation logged in that test’s record).

What does this workflow cost?

Almost nothing in generation fees: the reference is made once, and each generation simply cites one extra image. The real cost is verification time — every output gets frame-checked by a human, and that hour is not negotiable.

Item Notes Cost
Building the reference Crop + resize from the locked art, one-time About 20 minutes
A/B validation (2026-10-14) The same shot generated three ways (v1 / v2 / v3) 3 generations
Frame-by-frame QA Stills and video checked by eye, never outsourced to a metric About 1 hour per episode
The workflow itself One extra reference image per generation Effectively zero — what it saves is whole-shot redos

The full season-one production ledger — generation counts, subscriptions, the complete time account — is a separate post, and it will be published here in full.

Failures and limits: where the lock doesn’t hold

House rule: this section is never skipped.

  • This workflow exists because of a failure. When the production workflow was locked on 2026-10-02, character consistency was ranked above rain rules, weapon rules and everything else — because the drift across the first three episodes was the most expensive lesson we had.
  • It locks faces, not costumes. Wardrobe detail still has to be checked shot by shot against the character bible’s four locked outfits. The reference can’t help there.
  • Extreme angles still drift. Full profile and heavy occlusion carry visibly higher drift risk. The only defense is manual frame-checking — this cannot be fully automated.
  • The model “beautifies” the shading. In the v3 test, the face shading still came out slightly smoother than the flat matte color blocks of the reference. The model quietly polishes faces; we have not fully solved this, and it is recorded here as-is.
  • Camera moves don’t obey pixel-perfectly. First/last-frame generation treats the frames as references, not gospel: in testing, a locked-off long shot came back with a slight pull-back. If a shot needs a strict static frame, budget for a redo.
VERDICT

Recommended, with conditions — the face-reference workflow works: across 8 episodes of Night Huntress, the lead kept one face, not by luck but because every generation cited the same reference. Two conditions: (1) the reference must be cropped from the locked character art, never chained from generated frames; (2) human frame-checking stays in the loop, because extreme angles will still drift. If you are producing a multi-episode AI series, this is the first piece of infrastructure to build.

Sources and further reading

  • Night Huntress season-one production records and workflow document (DC Studio internal, locked 2026-10-02): source of the consistency rules, the reference spec and the QA method.
  • Full-length pipeline test log (2026-10-14): source of the test-shot02 v1 / v2 / v3 comparison results.
  • Character bible for Night Huntress: Li Ye’er’s four locked outfits and original portrait art.

FAQ

What image should I use to build a face reference?

A face close-up cropped from the character’s locked, final art. The source matters more than the crop: never chain a reference from an AI-generated frame, because each generation compounds the error until the face is no longer the same person.

Does every shot need the reference attached?

Yes. On Night Huntress, every generation featuring the lead — stills and video alike — cites the same face close-up, and the prompt states that her features follow it. No exception shots.

Does a face reference guarantee the face never changes?

No. It cuts drift dramatically, but full profile and heavily occluded shots remain risky. Our practice is to frame-check every output by eye and redo any shot where the face has moved.

Is the method for video the same as for stills?

Not quite. Stills just cite the face reference. For video, shots with a face toward camera work best in “first frame + face reference” mode instead of the usual first/last-frame pair — and we add a separate style reference to keep the brushwork consistent in motion.

Is this workflow expensive to run?

No. The reference is a one-time, roughly 20-minute job, and citing it in each generation costs nothing extra. The real cost is human QA — about an hour per episode — which we consider non-negotiable.

🧰 Tools used in this post

Full reviews, pricing and alternatives for the image and video generators live in the DC Affiliate review library — hands-on, with costs, like everything here.

Browse the tool reviews on DC Affiliate →

🎬 Want your own AI short-film series?

This pipeline is available as production and consulting at David Chen Studio: from character bible to a delivered season, built to the Night Huntress standard.

Talk to David Chen Studio →

The Fancy Rabbit field notes, once a month

The distilled version of this site’s hands-on tests — plus the cost figures and wreck reports that never make the posts.

One shared list across all three DC Studio sites · Unsubscribe anytime · We never sell your data

Related reading