Why AI Drama Characters Change Between Shots

Updated August 19, 2026 · Written by the AI Drama Reviews editorial team · 31 min read

DramaBoxEditor's pick — biggest short-drama library, lowest price to go unlimited 9.8Excellent

• Thousands of series, roughly 200 new titles every month
• Free episodes daily, plus ad-unlocks — no card needed to start
• Unlimited plans from about $5.99 a week, the lowest entry price we found
• Runs on Android, iOS and in a desktop browser

Go to DramaBoxAffiliate link · free to download · in-app purchases available

Affiliate link. We may earn a commission if you install through it, at no extra cost to you — it never changes our scores or placements. How we test · Disclaimer

Why do AI drama characters change between shots?

The complaint is right and the usual explanation for it is wrong. Viewers who say “the actress changes face halfway through episode 7” are describing something real and reproducible. What they are usually told — that the studio was lazy, or cheap, or rushed — does not survive contact with how the footage is made.

A conventional drama holds a character consistent by the crudest possible method: the same human being stands in front of the camera on Tuesday and again on Thursday. Continuity errors in live action are failures of record-keeping, and a continuity supervisor with a clipboard largely solves them. Nothing in that model transfers. In an AI drama there is no person, no wardrobe rail and no set. There is a description, and a system that redraws the character from that description every time it is asked.

So the honest framing is this: consistency in generated video is not preserved, it is reconstructed. Every shot is an attempt to hit the same target again from scratch. The attempts land close. They do not land identically, and the gap between “close” and “identical” is the thing your eye catches at the cut.

That is why more money helps and does not fix it. Budget buys more attempts, better references and more human hours spent throwing attempts away. It does not buy a model that remembers.

What you are actually seeing

The artefacts viewers report cluster into about ten recurring failure modes, and each one enters the finished episode at a specific point in the production pipeline. Knowing which stage produced which artefact is the difference between a useful complaint and a vague one.

Read the pipeline above from the point of view of a defect rather than a task. Stage 4, shot generation, is where identity drift, deforming hands, pseudo-text and physics violations are born, because that is the only stage that invents pixels. Stage 3, the character bible, is where drift is either constrained or left unconstrained. Stage 5 introduces the lip-sync problem, because in most workflows the voice arrives after the picture. Stage 6 is where all of it is either repaired, hidden or shipped. Stage 7, localisation, can reopen a lip-sync gap that stage 6 had closed, because a dub is a new audio track laid over mouth movements generated for a different language.

Why does the model forget between generations?

A video generation model does not hold a character. It holds a way of turning a description into pixels, and it starts from nothing every time you ask. This single fact explains almost every artefact on this page, and it is the one thing most coverage of AI dramas leaves out.

What actually happens when a shot is generated

A producer supplies a prompt, usually one or more reference images, sometimes a first and last frame, and a seed. The model turns that input into a few seconds of video. When the clip is finished, the process ends. Nothing about the character is written back into the model. The next shot begins with a fresh initialisation, a fresh prompt and the same references, and produces a fresh interpretation.

The comparison people reach for is a human artist drawing the same character twice, and it is a reasonable one as far as it goes. A skilled illustrator drawing a character for the two-hundredth time is working from an internal model built over months. The generation system has no such accumulation. Every shot is the first shot.

Why the references only narrow the target

Reference images make the model condition its output on an appearance. That is a real constraint and it does real work — it is the difference between a character who looks like a different person each shot and one who looks like the same person having a slightly odd day. But conditioning is statistical. The model is producing something highly probable given the reference, not retrieving a stored asset and re-rendering it. Highly probable leaves room for a jawline two percent wider and an eye colour half a shade off, which is exactly the range your visual system is tuned to detect in faces.

Why this is architectural rather than editorial

Because consistency is not stored anywhere, it has to be imposed from outside the model: by locking references, by keeping shots short enough that drift has less time to accumulate, by generating many takes and throwing most away, and by a human deciding which takes can sit next to each other. Every mitigation later on this page is a variation of that one idea. None of them is a fix, because none of them changes what the model does.

This is also why the production metric that matters is cost per accepted shot rather than cost per generation. A 90-second episode may require many minutes of raw generated material, and the discarded footage is not waste in the ordinary sense — it is the mechanism by which consistency is bought.

What is identity drift?

Identity drift is the failure mode viewers report most, because human faces are the thing human beings are best at reading. A one-millimetre change in the distance between the eyes is invisible on a chair and obvious on a person.

Drift shows up along several axes at once. Facial geometry moves — jaw width, nose bridge, the set of the eyes. Apparent age moves, so a character reads as twenty-five in one shot and thirty-two in the next. Body proportion moves, most visibly in shoulder width and height relative to furniture. Skin tone and lighting response move, which reads as a different person even when the geometry is stable.

Drift is worst in three situations, and they are predictable. It is worst on profile and three-quarter angles, because references are usually frontal and the model has less to anchor to. It is worst in low light, because the reference conditioning has fewer visible features to hold. And it is worst across long dialogue scenes, because those require many consecutive shots of the same face, which is precisely the arrangement that accumulates error.

The practical consequence for producers is that the shot list is written around the weakness. Coverage is planned to minimise consecutive close-ups of the same character, scenes are lit brightly and evenly, and profiles are avoided in favour of near-frontal framing. Watch enough AI drama and the flatness of the lighting stops looking like an aesthetic and starts looking like a workaround.

Why do props and costumes change between cuts?

Loss of object permanence is identity drift applied to everything that is not a face, and it is more common than face drift because nobody budgets review time for a teacup.

The mechanism is identical. A necklace exists in a shot because the model produced a plausible necklace, not because a necklace object was placed in a scene and tracked. Cut, generate again, and the necklace is re-invented — slightly different, or on the wrong side, or absent. The same applies to earrings, watches, tattoos, scars, buttons, bag straps, phone cases and the contents of a table.

Two patterns are worth naming because they are the easiest to spot deliberately. The first is the asymmetric accessory: an earring that appears in the left ear in one shot and both ears in the next, because symmetry is a strong prior in the training data and a single earring is not. The second is the disappearing hand prop: a character holds a glass, the shot cuts, and the glass is gone with no intervening action, because the second generation had no reason to produce it.

Why the set drifts too

Locations suffer the same problem at a slower rate. Furniture moves between shots of the same room, a doorway relocates, a window gains or loses a frame, a painting changes subject. Viewers forgive this far more readily than face drift, which is why producers spend their limited review budget on faces. It is nonetheless one of the more reliable signals if you are trying to identify AI production, and we rank it alongside the others in how to tell if a drama is AI-generated.

Physics and causality

Generated video learns what motion looks like, not what causes it, so it reproduces the appearance of physics without the constraints. The result is footage that is almost always plausible and occasionally impossible in ways that are hard to articulate but easy to feel.

Typical symptoms: objects that carry no apparent weight, so a heavy case swings like an empty one; cloth and hair that move convincingly for a second and then settle wrongly; liquid that pours without volume being conserved; a character who sits down into a chair whose seat is at the wrong height; contact between two bodies that is visually approximate, so a hand rests slightly inside a shoulder.

Causality fails at a larger scale. A door opens and the person who opened it is not positioned to have done so. A slap lands before the arm arrives. Someone reacts to an event a beat early. None of these is a rendering error; each is a consequence of a system that generates a sequence of frames that look right locally without maintaining a causal model of what is happening.

Editors handle this by cutting away from the moment of contact. Watch for it: AI dramas cut on impacts far more often than live action does, and the impact itself is frequently off screen or masked by an effect.

Why is the lip sync always off?

Lip sync fails in AI dramas because in most pipelines the mouth and the voice are made by different systems at different times, then reconciled in the edit. It is a workflow problem rather than a modelling limit, which is why it is one of the few artefacts that has genuinely improved.

The default sequence is generate video, then generate or record dialogue, then align. Alignment works when the generated mouth movement is generic enough to accept a range of audio, and stops working the moment the audio has a distinctive rhythm. Some newer tools change the sequence: the Sora 2 API produces synchronised audio with the video, priced at $0.10 per second at 720p portrait or landscape under the sora-2 model id, with Sora 2 Pro reported at $0.30 to $0.70 per second depending on resolution. Shots made that way start aligned.

Localisation reopens the gap immediately. A large share of what US viewers see on mixed-catalogue apps is a series produced in another language and dubbed into English by a voice model. The mouth movements belong to the original language and the audio does not. That is AI-localised drama rather than AI-generated drama — a distinction we set out in what are AI dramas — and it accounts for a good deal of the lip-sync complaint volume attributed to AI production generally.

Why does AI get hands and background text wrong?

Hands

Hands remain the single most reliable visual tell, because they combine every property that makes generation hard. They are small relative to the frame, so less model capacity is spent on them. They are highly articulated, with many joints moving independently. They self-occlude constantly, so the model must infer fingers it cannot see. And their correct configuration is strongly constrained — there are exactly five, they attach in a fixed order, and any deviation is instantly legible.

The characteristic failures are extra digits, fused digits, fingers that bend at impossible angles, and hands that change size relative to the face between shots. In a well-supervised pipeline these takes are discarded, which is expensive; in a volume pipeline they ship.

Background text

Written text in generated video usually comes out as pseudo-lettering: shapes with the visual rhythm of writing that are not actually characters. The model has learned what a shop sign, a newspaper or a phone screen looks like as texture. It has not learned to encode a string of characters, so it reproduces the texture.

This is why AI dramas are so often set in places without signage, why phone screens are shown from angles where nothing is readable, and why documents crucial to a plot are described in dialogue rather than shown. The workaround is avoidance. If you pause on any frame containing a sign, a book cover or a message thread, you will usually get an answer about production method within a second.

Why do all AI dramas use the same faces?

Unrelated AI dramas from unrelated producers frequently feel cast from the same small pool of actors, because models trained on overlapping data converge on the same faces. This is a category-level artefact and nothing downstream of generation can correct it.

Ask a video model for a wealthy male lead in his thirties, or a wronged heiress, or a brooding alpha, and you are asking for the centre of a learned distribution. Different systems trained on similar material have similar centres. Producers rarely fight this, because fighting it costs reference work and retakes for no gain in unlock rate. The result is a house look that spans the whole category.

It has a second-order effect on the drift problem. When two series share a face, drift in one becomes harder for a viewer to attribute: the character does not just look slightly different, she looks like someone from another show. This is most noticeable on AI-first catalogues such as FlexTV and on experimental slates such as MyMuse, where quality varies widely by title.

Why are AI drama shots so short?

Shot length in an AI drama is a model specification, not a directorial decision. Every model publishes a maximum clip duration, and those maxima — not taste, not pacing theory — are what set the rhythm of the finished episode.

Put those ceilings against the format. An episode runs 60 to 120 seconds. A scene is assembled from multiple short generations rather than one long render. Across the category, shots rarely exceed 10 to 12 seconds. Those three statements are the same statement viewed from three angles.

The economics compound the constraint. On Runway's plans — Free at $0 with 125 one-time credits and a watermark, Standard at $15 a month for 625 credits, Pro at $35 for 2,250 and Max at $95 for 9,500 — a single 10-second Gen-4.5 take consumes 120 credits. A Standard subscriber gets roughly five such takes a month before running out, and a production needs many takes per accepted shot. Google Flow's structure has the same shape: 1,000 credits a month on the Pro tier at $19.99 buys ten Veo 3.1 Quality generations, and monthly credits do not roll over. Google's own help pages are not internally consistent on Ultra pricing, which is worth flagging rather than smoothing over.

So the conclusion is blunt: when you watch an AI drama and notice that nothing is held for long, you are watching a specification, not a style. A director who wanted a 40-second unbroken two-shot could not have one at 8 seconds a generation, and could not afford the drift if the ceiling allowed it.

What can producers do about it?

Six countermeasures are in standard use, and every one of them reduces drift without eliminating it. That sentence is the honest summary of the entire mitigation stack, and any account that promises more is selling something.

What is a character bible?

A character bible is a locked reference set — canonical face images from several angles, wardrobe, hairstyle, accessories, height and build notes, plus location and prop references — that every generation in the season is pointed at. It is the highest-value artefact in an AI drama pipeline and the one most often under-built.

Under-building it is what produces the wardrobe-drift complaints specifically. If the bible says “dark business suit” and nothing more, every generation invents a lapel width, a button count and a shirt collar. If it specifies the garment down to the detail and supplies images of it, the variation collapses to something a viewer reads as lighting.

The limit is structural and worth stating plainly. A bible is a description handed to a system that approximates descriptions. It moves the output distribution; it does not select from a catalogue. Producers who have industrialised this — and FlexTV's move of North American production to AI, reported by MIT Technology Review on May 15, 2026, is the clearest documented case, with costs down 80 to 90 percent, schedules cut from three or four months to under one and teams of about ten people — still budget for retakes. That is a vendor describing its own results, and the retakes are the part that does not appear in the headline saving.

The bible also degrades across a season in a way nobody plans for. A hundred episodes are generated over weeks by rotating operators, each interpreting the same bible slightly differently in their prompts. The bible is stable; its application is not.

Reference images and frame control

Reference conditioning is the strongest per-shot lever available, and modern tooling has made it substantially better without making it sufficient. Google's Veo 3.1 supports reference images, input video, first and last frames, camera controls, scene extension, outpainting and object insert or remove, and applies a SynthID watermark to output across 1080p and 4K workflows. That is a consistency toolkit, and it is the direction the whole field is moving.

What first-and-last-frame control buys

If you can specify the frame a clip starts on and the frame it ends on, you can chain generations: the last frame of shot A becomes the first frame of shot B, and the seam is invisible. This genuinely fixes a class of cut-to-cut drift, and it is why chained sequences in recent titles look markedly more stable than they did a year ago.

What it does not do is constrain the interior of the clip, and it does not help when the next shot is a different angle in a different location three scenes later. Most drift viewers complain about is exactly that case.

Where reference conditioning breaks

Three situations defeat it consistently: extreme angles not covered by the reference set, lighting far from the reference conditions, and any shot where the character is doing something that changes their silhouette — running, fighting, being carried. Action beats are therefore where identity drift is worst, which is unfortunate given the genres this format favours.

Fixed seeds, and the myth around them

A seed initialises the random process inside a generation, and reusing it makes an identical input produce an identical output. That is all it does. It is a reproducibility tool, and it has been widely misrepresented online as a consistency tool.

The useful case is real. When you are iterating on one shot — adjusting a prompt, testing a reference — a fixed seed isolates your change, so you can see what your edit did rather than what the randomness did. Any serious pipeline uses seeds for this.

The misunderstanding is that reusing a seed across shots keeps the character the same. It does not. Change the prompt, the angle, the reference or the duration and the same seed lands somewhere else entirely, because the seed is the starting point of a process whose trajectory the inputs determine. Two shots of the same character in the same season are never identical inputs. Seeds contribute nothing to the problem this page is about.

Editing, human review, and the line that breaks first

The edit is where an AI drama is either rescued or shipped broken, and it is the only stage that catches everything. Continuity editing in this context is not assembly — it is triage.

The editor's toolkit is narrow but effective. Discard takes with visible drift and use another. Shorten a shot so the drift has less time on screen. Reorder coverage so two drifting shots are separated by a third party's reaction. Cut away from contact moments where physics fails. Add motion blur, effects or a lighting change over a seam. Re-time audio against a mouth that does not match. Insert a cutaway to avoid showing a prop that vanished.

All of that costs hours, and hours are the one input AI production did not make cheap. The 80 to 90 percent cost reduction FlexTV reported is a reduction in the cost of producing footage, not in the cost of judging it. Independent creator Jiang Lan's accounting of roughly $280 per finished minute — about $28,000 for twelve eight-minute episodes, reported by TVBS on August 7, 2026 — is a figure that includes this labour. Heavily optimised Chinese pipelines reported at around $30 per finished minute are getting there by industrialising the review step, not by skipping it, and that figure is not a typical US result.

What happens when volume production meets an unsolved problem

The review line is the first thing volume breaks. DataEye counted roughly 221,900 AI dramas published on Douyin in the first half of 2026 — about 1,200 a day — of which roughly 1,055, or 0.47%, passed 100 million views. At that output, per-shot human review of everything is arithmetically impossible, and the economics do not reward it: when 99.5% of titles will not find an audience, spending review hours on all of them is a losing bet.

The consequence for viewers is a quality distribution rather than a quality level. A small number of titles get real supervision and hold together; the long tail gets generated, lightly assembled and published. Complaints about AI drama consistency are therefore substantially complaints about the tail, and the tail is where the volume is. That is also why no useful generalisation exists about “how good AI dramas look” — and why no standardised independent quality benchmark ranks AI-drama seasons at all.

What the research literature actually says

Three published surveys make the same point in academic language: maintaining a coherent world across time is the open problem in video generation, and it is not close to closed. This matters for a reader because it settles the question of blame. The complaint is not about one studio's standards.

Read together, the three describe a field that has identified the problem precisely, agrees on its importance and has not solved it. Note the dates: the most recent of the three is three months old as of August 19, 2026, and it is a survey of open problems rather than an announcement of a solution.

Why does genre decide how bad the drift looks?

Fantasy, xianxia, werewolf and superpower stories survive drift; contemporary human drama does not. This is the most practical fact on the page, because it predicts which AI titles will be watchable before you press play.

Why stylised genres absorb the errors

Four reasons, and they compound. Stylisation gives the eye fewer fixed reference points — if the whole image is heightened, a face that shifts two percent reads as lighting or as art direction. Transformation is often part of the premise, so a character who changes is a plot event rather than an error; a werewolf story is structurally immune to the complaint this page is about. Effects that would consume a live-action microdrama's entire budget are close to free in generation, so the genre gets richer rather than poorer. And these genres tolerate rapid cutting, which is the format's only real defence against drift.

Why contemporary drama is the hardest case

Contemporary human drama depends on sustained micro-expression: a face holding an emotion across a long take while the other actor responds. That requires exactly what the models cannot deliver — a stable identity held for tens of seconds while it changes subtly and intentionally. It also puts the camera in ordinary rooms full of legible text, hands doing ordinary things with props, and phones being used on screen. Every weakness at once.

The audience evidence lines up with the technical explanation. Reviewer Jenny Cooper of Vertical Drama Love reported on June 24, 2026 that she watched 25 AI titles and finished three, concluding that story remained decisive and that fantasy was the strongest category. That is one reviewer's log, not a dataset, but it agrees with what the failure modes predict. We break the genres down individually in the AI drama genres guide.

Will it be solved?

The direction of travel is clear, the recent progress is real, and no source we hold claims the problem is solved. We are not going to put a date on it.

What can be said with a source attached: ByteDance released Seedance 2.0 in February 2026 and it is reported as a notable step up in generation quality. Google's Veo 3.1 shipped a set of features that are consistency tools by design — reference images, input video, first and last frame control, camera controls, scene extension, outpainting and object insert or remove. Runway released Gen-4.5 on December 1, 2025. Luma announced Ray3.2 on June 9, 2026. Sora 2 generates synchronised audio. Every one of those is a narrowing of the gap.

What cannot be said: that any of them removed the constraint. The ACM survey devoted to spatiotemporal consistency was published on May 18, 2026 — three months after Seedance 2.0 — and it is a survey of an open problem. Nothing in our fact base describes a model that maintains persistent character state across separate generations, and until something does, consistency remains a thing imposed from outside by references and human supervision.

There is a second-order point that is easy to miss. Even if generation quality keeps improving, the volume economics push the other way. When 1,200 titles a day are being published and 0.47% find an audience, better tools tend to produce more output at the same supervision level rather than the same output at a higher one. The average AI drama a US viewer encounters may not improve as fast as the models do.

Our position, stated so it can be checked: as of August 19, 2026 the consistency problem in generated video is unsolved, actively researched, materially improved on a two-year view, and mitigated in production by methods that reduce it without removing it. If that changes we will date the change on this page. Our standards for what counts as evidence are on the how we test page.

Frequently asked questions

The bottom line

Characters change between shots because each shot is a separate act of generation with no memory of the last one, and because the models cap out at roughly 8 to 12 seconds, so an episode is a chain of a dozen independent attempts to hit the same target. Consistency is not stored in the system. It is reimposed from outside, imperfectly, by references, a locked character bible and a human deciding which takes can sit next to each other.

That reframes the complaint rather than dismissing it. A viewer who notices drift is not being fussy and is not watching a badly made example of a well-solved craft. They are noticing the field's central open technical problem, which an ACM survey published on May 18, 2026 is entirely devoted to and which work on OpenReview in March 2026 and arXiv in October 2025 approaches from two other angles. Better funding buys more discarded takes and more review hours. It does not buy a model that remembers.

Practically, three things follow. Choose stylised genres if the artefacts bother you, because fantasy, xianxia and werewolf stories absorb what contemporary drama exposes. Expect quality to vary far more between titles on one platform than between platforms, because the variable is review hours per episode and that is decided title by title. And treat any claim that a new model has solved character consistency as a claim requiring a source, since as of August 19, 2026 no source we hold makes it.

If you want the production side of this in full, our how AI dramas are made guide walks the pipeline stage by stage, and every app we have tested against these criteria is in the reviews index.

DramaBoxEditor's pick — biggest short-drama library, lowest price to go unlimited 9.8Excellent

• Thousands of series, roughly 200 new titles every month
• Free episodes daily, plus ad-unlocks — no card needed to start
• Unlimited plans from about $5.99 a week, the lowest entry price we found
• Runs on Android, iOS and in a desktop browser

Go to DramaBoxAffiliate link · free to download · in-app purchases available

Affiliate link. We may earn a commission if you install through it, at no extra cost to you — it never changes our scores or placements. How we test · Disclaimer

Read next

Sources

• ACM — A Survey: Spatiotemporal Consistency in Video Generation, May 18, 2026. The survey treating consistency in space and time as the field's central open problem; the basis for this page's claim that drift is a research limitation rather than a production standard.
• OpenReview — work on narrative and temporal consistency in long video generation, March 26, 2026. Source for the distinction between short-clip consistency and story-level coherence across long sequences.
• arXiv:2510.04999 — Bridging Text and Video Generation: A Survey, October 6, 2025. Source for the gap between what a text description specifies and what a model produces, which is why a character bible is a specification rather than a control.
• Model specifications as published by their vendors — ByteDance Jimeng (~12 seconds per generation, 30 with motion templates); Kling (up to about 3 minutes); Sora 2 and Runway Gen-4.5 (up to 1080p); Google Veo 3.1 (Lite and Fast at 4, 6 or 8 seconds, Quality at 8 seconds; reference images, input video, first and last frames, camera controls, scene extension, outpainting, object insert and remove; SynthID watermark). Seedance 2.0 released February 2026; Runway Gen-4.5 released December 1, 2025; Luma Ray3.2 announced June 9, 2026.
• Published pricing pages and reported rates — Sora 2 API at $0.10 per second at 720p portrait or landscape under model id sora-2, Sora 2 Pro reported at $0.30–$0.70 per second; Runway Free, Standard $15/mo, Pro $35/mo and Max $95/mo, with Gen-4.5 at 12 credits per second; Google Flow free 50 credits/day, Plus 200/month, Pro 1,000/month at $19.99, Ultra tiers listed at $100 and $200. Google's own pages are not internally consistent on Ultra pricing, and several tool prices in circulation come from secondary sources rather than retrieved official pages.
• MIT Technology Review (May 15, 2026) — FlexTV vice-president Tang Tang on the move of North American production to AI: costs down 80–90%, schedules from three or four months to under one, teams of about ten. A vendor describing its own results.
• DataEye (2026) — approximately 221,900 AI dramas published on Douyin in H1 2026, roughly 1,200 a day, of which about 1,055 (0.47%) passed 100 million views; ~470 AI titles a day in January 2026 via MIT Technology Review. China Netcasting Services Association via Global Times — more than 95% of new Chinese microdramas AI-generated in Q1 2026.
• TVBS (August 7, 2026) — independent creator Jiang Lan's accounting of roughly $280 per finished minute and about $28,000 for twelve eight-minute episodes. NOW News (December 15, 2025) — 60–80 episode AI microdramas produced for $3,000–$14,000. Reported optimised-pipeline figures of ~$30 per finished minute describe a Chinese workflow and are not a typical US result.
• Vertical Drama Love (June 24, 2026) — reviewer Jenny Cooper watched 25 AI titles and finished three, concluding story remained decisive and fantasy was the strongest category. PetaPixel and Today (August 2026) — director Qingge Gao on microdrama viewers who “can easily watch AI minidramas… most of them don't care.” Reddit threads on an AI mythological serial. All are stated opinions or single-source observations, not measured data.
• Format specifications reported across the category — 9:16 vertical, episodes of 60–120 seconds, seasons of 40–100+ episodes, scenes assembled from multiple short generations, shots rarely exceeding 10–12 seconds, and cost per accepted shot rather than cost per generation as the operative production metric.
• Declared gap — no standardised independent quality benchmark ranks AI-drama seasons, and no authoritative dataset shows what percentage of any app's catalogue is AI-generated. Statements on this page about quality distribution are inferences from production economics, and are labelled as such.

Affiliate disclosure. Some outbound links on this page, including the DramaBox links, are affiliate links. If you install through one we may receive a commission at no additional cost to you, and it does not change what we write — this page applies the same technical criticism to DramaBox's AI shelf as to every other catalogue. Accuracy notice. Model specifications and prices in this category change without notice, several figures here come from secondary sources or from vendors describing their own results, and we label those rather than presenting them as verified. Nothing on this page should be read as a claim that any model has solved character consistency. Verified August 19, 2026.