5.1 The Reference System: The Sheet and Its Role
Part IV produced an identity: a trained adapter and a plate. Part V turns it into a production system that yields dozens of consistent frames a week.
The first component is the The versioned plate of the persona's face and body that every generation is scored against. — and the single most common mistake in this part of the work is treating it as a starting point. It is not. The sheet is an input to every generation, permanently. The moment it becomes a thing you used once, The slow migration of a face away from the reference plate, invisible in one frame and obvious across a year. starts, and by the time you notice it you have three weeks of published content with three slightly different faces.
What the sheet contains
A 3×3 grid of nine frames, generated from the trained adapter and the plate, at high resolution, against a neutral background, with even light and no styling:
| Column A | Column B | Column C | |
|---|---|---|---|
| Row 1 | Medium close-up, neutral | Medium shot | Wide shot |
| Row 2 | Full body | High angle | Low angle |
| Row 3 | Left profile | Three-quarter right | Back, head turned |
Nine frames, one identity. Then a second sheet for the body: the same person in the same unstyled clothing at full height, front and back, so clothing reference is separate from face reference.
Four requirements on the sheet itself:
High resolution, above your typical output. If you generate at 1024, the sheet is at 2048 or higher. The sheet is the thing other generations are pulled toward, and a soft reference produces soft results.
Neutral light and no styling. No makeup decisions, no wardrobe choices, no location. Those are 5.3's layer. A sheet with styling baked in drags that styling into every generation, which is the dataset contamination problem from 4.4 arriving through a different door.
Validated against 4.6. Score the sheet before you use it. A sheet built from a failing One saved state of a model or adapter, identified by a hash and recorded as provenance. distributes the failure through your entire pipeline.
Versioned with the checkpoint. Sheet v3 belongs to A small low-rank adapter, tens to a couple hundred megabytes, that nudges a base toward one face. v3 on base X. Mixing them is the
base-mismatch trap from 4.5.
Building the sheet: the procedure
Six steps, roughly two hours, and worth doing in one sitting so the light and the styling are consistent across all nine frames.
- Confirm the checkpoint. Score the adapter against the 4.6 core battery first. A sheet built on a failing checkpoint distributes the failure everywhere.
- Lock the conditions. One neutral background, one even light setup, no styling, no wardrobe decisions beyond plain clothing. Write the conditions down — they are part of the sheet's specification.
- Generate the nine. Each frame from the same adapter, the same conditions, at above-output resolution. Do not cherry-pick across sessions; a sheet assembled from different days and different prompts has inconsistent light.
- Reject and regenerate until all nine pass. Any frame that scores below 4 on facial structure (4.6) is regenerated, not kept. The sheet is the reference everything else is pulled toward; an average frame here propagates.
- Build the body sheet separately. Full height, front and back, plain clothing, same conditions. Proportions are scored independently (4.6) and need their own reference.
- Name, version, and archive the source prompts.
face-sheet-v3-loRA-v3-base-Xplus the prompts that produced it. If you ever need to rebuild the sheet, you rebuild it from the record rather than from memory.
Re-run the procedure when the adapter version changes, and not before. A sheet is expensive to build and cheap to keep.
Why it must be a permanent input
Three mechanisms, each of which removes a class of drift.
It anchors the distribution, not the average. A trained adapter already knows the face, but it also has a wide output distribution. The sheet narrows where in that distribution the model lands. Without it, each generation is a fresh sample; with it, each generation is pulled toward nine fixed points.
It fixes what text cannot describe. Prompt text can say "same face", and it changes nothing — there is no word for a specific person. A reference image is the only mechanism that carries the particular.
It makes a mistake checkable. With a sheet, "does this match?" has an answer: put the frame beside the sheet. Without it, you are comparing against memory, which is the scoring error from 4.6.
The reference image is a permanent input node, not a fragment of prompt text. If your workflow requires you to re-upload the reference for each generation, that is not a preference — it is a design flaw that will produce drift on the day you forget. Build the reference into the workflow itself (5.4).
The reference-conditioned method, honestly assessed
There is a widely-shared approach that uses a reference sheet as an in-context example to a general-purpose image model: supply the sheet as knowledge, plus a written description of the character, and ask for a new scene.
What it offers. No training, no adapter, usable within an hour. For a persona that has to exist today, this is the fastest The numbered folder convention such as 00-bible/ and 40-output/ that keeps every asset addressable. from zero to publishable, and chapter 4.2 treats it as the entry rung of the The identity techniques ordered from cheapest and weakest to most expensive and most portable..
What it costs. [UNVERIFIED] The approach is inconsistent in ways that are difficult to predict. Operators report that it holds for a while and then drifts — the face begins to migrate, particularly around the eyes and jaw, and the drift is not clearly correlated with anything in the prompt. It also tends to reproduce the style of the reference sheet, since a general model treats the supplied images as a stylistic signal as well as an identity signal.
The honest framing: the method is a valid way to start and a poor way to continue. Use it to validate that the persona works before you spend a weekend training (4.7). Do not build a year of content on it and then discover the drift, because by then you have published the drift.
If you do use it, two mitigations: keep the sheet's styling completely neutral so there is no style to leak, and re-check identity weekly rather than quarterly.
How the sheet is used in a generation
The order of operations, which matches 4.7's two-layer architecture:
- Identity layer: the adapter, plus the sheet as Instruction-based edit models and adapters that hold the subject while the scene changes.. Frozen settings.
- Scene layer: the prompt describes location, wardrobe, light, action. It never describes the face.
- Composition: framing chosen from the kit's rules (2.5).
The rule that prevents most identity failures: the prompt describes everything except the person. The moment a prompt contains facial description, two sources of identity are competing — the reference and the text — and the text wins more often than you expect. This is the same Naming in text what an image contains; what you name is what separates from the trigger word. principle as 4.5, applied at generation time.
Reference for body, wardrobe, and location
The face is one reference layer; there are three more, and they belong in the same folder with the same discipline.
Body reference. A full-height sheet in unstyled clothing, front and back. Face consistency does not imply body consistency — proportions drift independently, which is why 4.6 scores them separately.
Wardrobe reference. One image per outfit, on the persona, front view. The wardrobe is a reusable asset, not a per-post decision, and its consistency is not the adapter's job.
Location reference. One clean plate per recurring location: the kitchen, the street, the cafe, empty of people. Locations are the cheapest consistency win available, because a location can be generated once and reused for years while a face has to be held.
Chapter 5.3 builds all three. The point here is that they are references in the same sense the sheet is — permanent inputs, versioned, and not re-improvised per post.
Where this goes wrong
Using the sheet once. Drift is now a matter of time.
A sheet with styling baked in. The style propagates into every generation.
A low-resolution sheet. Everything downstream inherits the softness.
Re-uploading the reference manually. It will be forgotten, and the forgotten generation is the one that publishes.
Describing the face in the prompt. Two identity sources competing; the wrong one wins.
No body or wardrobe reference. Face consistency without proportion consistency, which audiences notice even when they cannot name it.
Sheet and checkpoint versions mixed. The mismatch is subtle and it looks like drift.
The gate
- Do you have a 3×3 face sheet, at above-output resolution, neutral light, no styling?
- Is it scored against 4.6?
- Is the sheet a permanent input in the workflow, not a manual upload?
- Do body, wardrobe, and location references exist as separate layers?
- Does your prompt template describe everything except the person?
- Is the sheet versioned with the checkpoint it belongs to?
What you leave with
A 3×3 face sheet with four quality requirements, a separate body sheet and the two further reference layers, the mechanism argument for why the sheet must be a permanent input rather than a starting point, an honest assessment of the reference-conditioned method including what it costs and when to move past it, and the operating rule that carries most of this part: describe everything except the person.