Consistency is a pipeline

All read on 18 September 2026 from the vendors' documentation pages, and every one of them is dated on the control sheet because every one of them moves.

• The seed binds the starting noise, so the same brief at the same version gives the same picture. Midjourney's compatibility chart puts the worth of that in two words: "(99% identical)" on V8.1 and V8.2. Not exact. One asset in a hundred is not the one you approved, and across six assets and four revisions each, that is a number you will meet.
• The style reference --sref binds look and feel in proportion to --sw, which runs 0 to 1000 with a default of 100. The page says what it carries: style, "not objects or people". And a style code cannot be made from an uploaded image, so the approved mug is not a style; it is a subject, and this handle will not hold it.
• Subject continuity is the Edit Model on V8.1 and V8.2: up to four reference images, and it replaced Omni Reference and Character Reference in one release. On V7 the same job was Omni Reference, one image at 2x GPU cost with --ow 1 to 1000. Google caps references per model (up to fourteen on the lightest, six object and five character and three style on the Pro). OpenAI's own guide says the model "may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations". That sentence is the gap, documented by the vendor.
• The adapter is the only handle that can carry identity, and it is not a flag. It is a training job you run yourself, and the self-study card below is about deciding whether to.

Read the four labels in the figure again. Style. Noise. Up to four images. A training job. Identity, the thing the head of marketing asked for, is on none of them. That is why the answer is a process and not a parameter.

LoRA, per the Hugging Face Diffusers documentation, inserts "a smaller number of new weights" into the model and trains only those, producing adapters of "a few hundred MBs" that target the attention projections through PEFT. It is the one handle that binds identity, because it changes the model rather than persuading it. Three facts from the same page decide whether Daybreak should touch it.

• It is a training job with a cost. The reference run in the docs is "~5 hours on a 2080 Ti GPU with 11GB of VRAM", for one adapter, on one base model, before anyone has judged whether the result looks like the brand.
• It is a maintenance burden across base-model versions. An adapter is trained against one base. When the base moves, and the defaults have moved three times in fifteen months, the adapter is retrained or it drifts, and nobody will tell you which.
• It is unsupported by the vendors whose models most people use. The Diffusers page itself is flagged "experimental and the API may change in the future". No hosted vendor in the table below offers it as a control. It is yours to run, yours to host, yours to keep working.

So the adapter is the decision "when nothing else will do": a recurring character on a hundred assets a year, a mascot that must be that mascot, a product whose shape is the brand. For six autumn roast illustrations and a mug, the Edit Model and a named person comparing outputs is the proportionate answer, and it costs nothing to maintain. The adapter is on the control sheet as a line that says "not this job, and here is the threshold at which it would be".

No single control gives identity. The pipeline does, because each step holds the one property its handle can hold and hands the rest downstream. The version is locked and dated, from session 1, so the encoder under every asset is the same encoder. The kit is named and weighted, from session 2, so style is a file with a number and not an opinion. The composition reference holds the zone, from session 3, so the picture leaves room for the type. The type never enters, from session 4, so nothing the model cannot do is asked of it. The seed is recorded with its ninety-nine written next to it, so a match is expected and a drift is not a surprise.

Then the step people leave off: a person compares every output to the six sentences and rejects on sight. Not "does it look nice", but "is the light from the left, is the mug the approved mug, is the grain the grain the sentence describes". That person is a step in the pipeline, not a check after it, and is on the control sheet by name. A pipeline whose last box says "someone" has no last box.

Three defaults in fifteen months. The version will move under a live set, probably mid-campaign, and the pipeline needs a rule for it before it happens rather than an argument after. The rule has three parts.

• When the default changes, the pipeline is re-run once on the kit. Same brief, same references, same recorded seed, ten runs. The outputs are compared to the six sentences, in writing, not to memory of what last month's tiles looked like. Memory is the thing that drifts.
• If the kit holds, the control sheet gets a new date and nothing else changes. The set continues on the new default from the next asset onward.
• If the kit does not hold, the kit is re-cut: the reference that carried the property that moved is replaced or its note is sharpened, and the ten runs go again. The earlier assets are not regenerated. They were approved under a dated version, they are on record, and a set is allowed to contain assets made under two versions if a person compared both against the same six sentences.

The US Copyright Office published Part 2 of its report on copyright and artificial intelligence on 29 January 2025. Two lines from the PDF are the whole card. First: "prompts do not alone provide sufficient control". A prompt is a persuading layer, and the Office reached the same conclusion session 1 did, from its own side. Second: human authors hold copyright in "the creative selection, coordination, or arrangement of material in the outputs, or creative modifications of the outputs".

Read that second line against the pipeline. A composition reference that holds the zone is a selection. A style kit at a recorded weight, four files with names, is coordination. A type layer set by hand over a picture layer with the lettering refused is both an arrangement and a modification. The pipeline taught in sessions 1 to 4 is exactly the category the Office describes, and the control sheet is the record of it. The pure prompt, one sentence and one PNG, is exactly the category it excludes.

That is the one fact the pipeline needs from that ground. Copyright beyond this line, indemnities, the court cases and provenance as a signature are learn-ai-design a4. Disclosure duties and the EU dates are learn-ai-media 06. Both pages are current and both are owned by someone whose job it is to keep them so.

Look at what is on the sheet by the end of tonight. A dated version. A named kit with a weight. A recorded seed with its ceiling written next to it. A placement line for the zone. A type layer with the exact string and who approved it. A named person who compared ten runs against six written sentences and rejected three. That is a documented sequence of human choices, each one dated, each one pointing at a file or a page.

A saved prompt is not that. It is one input to a persuading layer, and it records nothing about what was selected, arranged or modified afterwards, because nothing was. The same sheet that makes the set consistent is the sheet that shows a person made it. Consistency and authorship turn out to be the same paperwork, which is the best argument there is for keeping it.