Four consistency mechanisms and when each fails

When a batch of shots doesn't match, the reflex is to reroll the offending one. That works often enough to become a habit and it's still the wrong instinct, because a reroll teaches you nothing — the same shot fails the same way the second time about as often as it doesn't, and you've spent credits finding that out. Four mechanisms hold a look in place across generations. They lock different things, and each has a failure signature specific enough to name from the output. Diagnose first; the reroll is what you do after you know which of the four let go.

What each one locks

Mechanism	Locks	Doesn't lock	Setup
Named reusable assets	A specific entity — this character, this product, this location — by key across every scene in a build	Anything not registered as an asset; the grade	One upload per entity, once
Reference stacking	Several traits at once in a single call — face, product, set, wardrobe	Anything past the endpoint's reference cap; continuity between calls	Per-call attachment
Style references	Palette, grade, lens character, the "same photographer" quality	Identity, geometry, subject detail	Per-call, or saved as a preset
Fine-tune / LoRA	The subject into the weights, so it needs no attachment at all	Anything outside the training distribution	Curated dataset, front-loaded

That table is the whole argument compressed: these are not four price points on one product. They lock four genuinely different things, and the wrong one produces output that is confidently, consistently wrong rather than randomly wrong.

The four failure signatures

Named assets fail by coverage and by omission. A named asset is an entity registered once with reference images and then referenced by key. In Versely that's request_asset_upload opening the Asset Builder, or create_workflow_asset if the images are already hosted, bound to a workflow. The published workflow library shows the shape: 17 workflows carry 65 registered assets between them across 166 scenes — 34 characters, 18 scenes, 7 locations, 3 products, 2 props and 1 scene reference.

It fails two ways, and they look different. Coverage failure: the asset is right in every shot that resembles its reference images and wrong in the ones that don't. Three front-facing portraits will not carry a shot from behind, and the model fills the gap by inventing. The signature is that correctness tracks camera angle. Omission failure: one scene didn't list the asset in its reference keys, so that scene silently ran on prompt alone. The signature here is unmistakable once you know to look for it — every scene is right except one, and the wrong one is wrong in every respect at once rather than drifting slightly. Drift is gradual. Omission is a cliff.

Reference stacking fails by dilution and by cap. Stacking means several anchors in one call: reference_images on generate_image_from_image, or reference_image_urls and reference_video_urls on the video tools. Caps vary a lot across the catalog — image endpoints run from a single reference up to 16, and on the video side Wan 2.7 Reference to Video takes up to five reference images and/or videos. Reference stacking covers the endpoint-by-endpoint picture.

Dilution is the characteristic failure: attach six references and the model averages them rather than honouring each. The signature is that everything is slightly off and nothing is badly off — no single element you can point at, just a general sense that the shot isn't the brief. The trait you cared most about is usually the one that lost, because you attached it alongside five competitors. Cap failure is dumber: pass more references than the endpoint accepts and the extras are dropped, usually without a visible complaint, so a specific reference appears to have had no effect at all. get_model_input_schema returns the exact accepted fields and bounds for an endpoint, which settles this in one call rather than three rerolls.

Style references fail by scope. A style preset or style reference locks the grade — palette, contrast, lens character, film stock feel. Ninety-one models in the catalog declare style support. What it does not lock is who is in the frame. The signature is exact and easy to recognise once named: every shot looks like it came from the same photographer and a different person. If your batch has coherent colour and a subject whose face changed, you did not have a style problem. You had an identity problem and you solved a style one.

Fine-tunes fail by memorising the room, and by going stale. A fine-tune or LoRA moves the subject into the weights so nothing needs attaching. Its failure is a training-data failure wearing a generation costume: a dataset shot in one room on one afternoon teaches the room as hard as it teaches the subject. The signature is a subject that is perfect and immovable — always the same light, the same three-quarter angle, resisting every attempt to place it somewhere new. The second signature is temporal: the checkpoint doesn't know about the packaging redesign, so it keeps producing last season's product with total confidence. Style references vs fine-tunes covers where the crossover sits.

Be accurate about scope before planning around it: on Versely the agent's fine-tune tools are voice-side and read-only — list_finetunes, get_finetune_status and list_finetune_voices enumerate voice fine-tunes on an account and report whether one is done. There is no image fine-tune training tool in the agent today, so treat that column as the general mechanism rather than a workflow you can start from a chat message.

Diagnosing from the output

Run this against a batch before rerolling anything.

What you see	Mechanism at fault	Fix
One scene wrong in every respect, the rest fine	Named asset not referenced in that scene	Add the asset key to that scene
Subject correct head-on, wrong from other angles	Asset reference set lacks coverage	Add references at the missing angles
Everything slightly off, nothing badly off	Reference dilution	Cut the stack to the two or three that matter
One reference had no visible effect	Endpoint reference cap exceeded	Check the schema; resubmit under the cap
Same look, different person	Style reference used where identity was needed	Add an identity anchor; keep the style
Same person, different grade every shot	No style mechanism at all	Add a style preset or reference
Subject perfect, always in the same room	Fine-tune memorised its capture conditions	Vary angle, light and background in the dataset
Subject correct but out of date	Stale checkpoint	Retrain, or fall back to references

Each row implies a different action and none of them is "reroll." Consistency collapse breaks the same territory down by the four kinds of drift — identity, wardrobe, set and lighting — the complementary view if you'd rather start from the symptom's category than its mechanism.

Combining them without cancelling them out

The mechanisms compose, but only in one direction. Identity anchors and style anchors coexist happily: a named asset holding the character and a style reference holding the grade solve different problems and don't compete. Two identity anchors do compete — a named asset plus a loose stack of extra portraits of the same person is the dilution failure with extra steps, because you've given the model several slightly different accounts of one face and asked it to pick.

A working order:

1. Register the entities that must not change. Characters, hero products, recurring locations. Registering an asset isn't a generation, so it doesn't draw generation credits — the generations that use it still do.
2. Attach a style anchor separately, and keep it separate. Grade is not identity.
3. Keep the stack small. Two or three deliberate references beat six hopeful ones; the cap is a ceiling, not a target.
4. Let the scene writer restate traits rather than restating them yourself. Versely's scene director reads each asset's reference images inline and writes wardrobe, palette and lighting back into every scene prompt that references it.
5. Reserve fine-tuning for subjects with a horizon measured in months. A campaign of twenty assets never earns back a curated dataset.

FAQ

If I have a named asset, do I still need reference images in the prompt?

The asset is reference images — that's what registering it does. Attaching more of the same subject on top of it usually costs you rather than helping, because you've introduced competing accounts of one identity. Add references for things the asset doesn't cover, not for the thing it does.

How many references is too many?

Below the endpoint's cap, and below the point where they start disagreeing. Two or three references of one subject from genuinely different angles is stronger than six near-duplicates, and six references of six different things is a dilution problem regardless of what the cap allows. If the trait you care most about is the one drifting, that's the signal to cut the stack.

Can a style reference fix a character who keeps changing?

No, and this is the most common misdiagnosis in the list. A style reference locks the look of the image, not the subject in it. The shots will get more coherent and the face will keep changing, which reads as improvement for about a day. Character consistency needs an identity anchor — an asset or a reference image of the person.

What do I do when the output is right but I can't tell which mechanism is holding it?

Remove one and regenerate. It's the only reliable way to find which anchor is load-bearing, and worth doing once on a build you'll repeat — the answer tells you what to protect when the brief changes. Do it on one scene, not the whole batch; reroll budgeting makes the case for small diagnostic runs.