Appendix A — Glossary

This appendix is the book's working vocabulary, grouped by the layer of work each term belongs to rather than alphabetized, so you can read the section matching the chapter you have open. Keep it beside the machine: the term you cannot name is usually the variable that broke the batch. Every entry is one sentence and a chapter reference, and the chapters carry the qualifications and the failure modes. The compliance section will age fastest. [AS OF 2026-10]

Convention: "Where:" lists the chapters where the term is used. Entries marked [FIELD] are standard field vocabulary the book covers at category level without using that exact label. The annotation is not an honesty marker; markers are defined in How to Read This Book.

Identity, canon, and characterization

Character bible — The six-part specification that makes a persona reproducible as a character across months. Where: 2.3, 2.7.

Canon — The ten facts a persona cannot break, each specific enough to check in ten seconds. Where: 2.3, 5.6.

Timeline register — The dated record of what the persona has already done or claimed, used to catch contradictions. Where: 2.4, 5.6.

Prop closet — The recurring places and objects around the persona; what turns generation from invention into retrieval. Where: 2.3, 5.3.

Reference sheet — The versioned plate of the persona's face and body that every generation is scored against. Where: 5.1, 4.6.

Identity drift — The slow migration of a face away from the reference plate, invisible in one frame and obvious across a year. Where: 4.1, 4.6.

Identity score — The 1–5 per-frame rating; nothing scoring 2 or below survives, whatever the mean says. Where: 4.6, 5.6.

Models, training, and identity adapters

Base model — The trained weights a persona's look comes from; its license governs any adapter trained on it. Where: 3.1, 4.5.

Checkpoint — One saved state of a model or adapter, identified by a hash and recorded as provenance. Where: 4.5, 11.5.

LoRA — A small low-rank adapter, tens to a couple hundred megabytes, that nudges a base toward one face. Where: 3.2, 4.5, 4.7.

LoRA rank and alpha — Rank sets the adapter's capacity; alpha scales how hard it pushes. Capacity, not quality. Where: 4.5.

DreamBooth — Fine-tuning more of the model on a handful of images: higher fidelity, gigabyte outputs, more overfitting. Where: 3.2.

Dataset curation — Designing rather than collecting the training set: score a generated pool for identity, then for coverage. Where: 4.4.

Captioning — Naming in text what an image contains; what you name is what separates from the trigger word. Where: 4.5.

Trigger word — The token activating the adapter, and the token that absorbs whatever the dataset wrongly associates with the face. Where: 4.5.

Regularization images — Plain images of the class an adapter teaches, included to hold the base model's general behavior steady. [FIELD] Where: 4.4, 4.5.

Overfitting and underfitting — Overfitting memorizes the reference set including its lighting; underfitting leaves a face too weakly learned to hold. Where: 3.2, 4.5.

Method ladder — The identity techniques ordered from cheapest and weakest to most expensive and most portable. Where: 4.2.

IPAdapter — Conditions generation on reference images, transferring style and composition as well as face; unreliable as identity alone. Where: 3.2, 4.3.

IP-Adapter FaceID — IPAdapter carrying an identity embedding extracted by a face recognition model, which sharpens retention. Where: 3.2, 4.3.

InstantID — A face embedding plus a landmark path, so one reference holds identity as pose and scene change. Where: 3.2, 4.3.

PuLID — Injects an identity embedding with contrastive alignment designed to leave the base aesthetic untouched; no training. Where: 3.2, 4.3.

ControlNet — A conditioning path constraining generation with a structural map such as pose, depth, or edges. [FIELD] Where: 4.3.

ReActor-style face swap — Replacing a face after generation: fast and cheap, weak across angles and motion. [FIELD] Where: 4.2, 4.7.

Expression echo — The adapter reproducing whichever expression or pose dominated its dataset; a concentration failure, not a prompt failure. Where: 4.3, 4.4, 5.6.

Latent space — The compressed representation the model denoises in, which is why img2img is approximate and inpainting is soft-edged. Where: 3.2.

VAE — The encoder-decoder pair between pixels and latent space; a mismatched one shows as color shift and washed shadows. Where: 3.2.

CLIP — A CLIP-style text encoder that responds best to comma-separated fragments; used by older model families. Where: 3.2.

CLIP Vision — The image-tower half of a CLIP-family encoder, conditioning on a reference image rather than on text. [FIELD] Where: 3.2, 4.3.

T5-class encoder — A long-context text encoder that handles full sentences and word relationships; write prose, not tag soup. Where: 3.2.

Sampling and inference settings

Seed — The random starting point; the same seed with the same settings reproduces the same image. Where: 3.2, 3.5.

Step count — How many denoising iterations run; full-quality bases converge near 20–30, distilled models near 4–8. Where: 3.2.

CFG scale — Classifier-free guidance: how hard the model is pushed toward the prompt; high values over-bake an identity LoRA. Where: 3.2.

Sampler — The algorithm walking from noise to image, such as Euler or a DPM++ variant. Where: 3.2.

Scheduler — The noise schedule the sampler follows, such as normal or karras. Where: 3.2.

Sampling preset — The frozen sampler, scheduler, steps, and CFG written into the persona's specification. Where: 3.2, 3.5.

Editing, conditioning, and post-processing

txt2img — Generation from text alone with no reference; where a persona starts and the wrong tool for maintaining one. Where: 3.1, 3.2.

img2img — Denoising a whole image from your input at a strength value, restyling a shot while keeping its framing. Where: 3.2.

Inpainting — Regenerating only a masked region; the surgical tool for a hand, an eye, or one bad detail. Where: 3.2, 5.5.

Outpainting — Extending an image beyond its original frame; the book composes wider at generation instead. [FIELD] Where: 5.5.

Reference conditioning — Instruction-based edit models and adapters that hold the subject while the scene changes. Where: 3.2, 5.1.

Detail pass — A face-detail operation run last at final resolution and low denoise; over-detailing reads as synthetic. Where: 5.5, 3.2.

ADetailer — A face-detail pass applied after upscaling; the book names the operation rather than this button. [FIELD] Where: 5.5.

Upscaling — Raising resolution on pixels rather than latents, so it runs after composition, identity, and repair. Where: 3.2, 5.5.

Rejection table — The twelve-defect diagnostic routing nine of twelve fixes back into the pipeline rather than into the frame. Where: 5.6.

Publishing queue — The four destinations decided at QA rather than later: publish, bank, retouch, or discard. Where: 5.6.

Video, motion, and audio

I2V (image-to-video) — Generating motion from a validated still; the still-first workflow is what makes keyframes portable. Where: 6.1, 6.2.

T2V (text-to-video) — Generating a clip from text with no reference frame; the weaker route for a persona. Where: 6.1.

Keyframe — The validated still a clip is generated from, carrying the identity the video model would otherwise sample. Where: 6.1, 6.2.

Start frame and end frame — The images a clip interpolates between; specifying both constrains motion and reduces drift. Where: 6.2, 6.4.

Motion prompt — The text describing movement alone, kept short when identity conditioning is heavy. Where: 6.2, 6.3.

Motion bucket — The parameter controlling how much movement a clip is asked for. [FIELD] Where: 6.2.

Temporal consistency — Whether appearance holds across the frames of one clip; the video form of drift. [FIELD] Where: 6.1, 6.3.

Flicker — Frame-to-frame instability in brightness, texture, or detail; a generator limitation, not a grading problem. [FIELD] Where: 6.1, 6.7.

Lip-sync — Matching synthesized speech to a face; priced per minute and the most retried line in the cost sheet. Where: 6.6.

Phoneme — A unit of sound a lip-sync model maps to mouth shapes; difficult ones triple the retries. Where: 6.6.

Talking head — A synced speaking clip; the format that moves a pipeline from Minimum to Working. Where: 6.6, 7.5.

Voice cloning — Producing new speech from a sample of a real voice; anchoring to a fixed artifact is why it beats voice design. Where: 6.5.

Voice design — Describing a voice and having a model create one with no source recording: no consent problem, less consistency. Where: 6.5.

TTS (text-to-speech) — Synthesis of speech from text; the cheapest layer in the stack at solo volume. Where: 6.5.

Zero-shot cloning — Cloning from a short sample with no fine-tuning, at the cost of emotion-to-emotion stability. [FIELD] Where: 6.5.

Dubbing vs regeneration — Dubbing replaces the audio track over existing visuals; regeneration re-produces the clip, which the mismatch rule prefers. Where: 6.4, 6.7.

Pipeline, automation, and production

Path — The numbered folder convention such as 00-bible/ and 40-output/ that keeps every asset addressable. Where: 3.6, 5.3.

Node graph — A saved generation workflow expressed as connected nodes, runnable from an API or script. Where: 7.2.

ComfyUI — The node-graph engine the book uses as a production engine, callable by an orchestrator. Where: 3.4, 7.2.

Workflow JSON — The serialized node graph exported from the editor, which is what the orchestrator executes. [FIELD] Where: 7.2.

Orchestrator — The automation layer that calls the engine, passes parameters, and files outputs; n8n here. Where: 7.2, 7.3.

Approval gate — The human check immediately before the irreversible step: publication, or a render you pay per second for. Where: 7.1, 7.3.

Smoke test — Three frames at production settings, checked for identity, before every large batch. Where: 3.4, 7.2.

Run id — The identifier given to one pipeline run, making a re-run idempotent instead of duplicating outputs. Where: 7.2, 7.3.

Prompt list — The maintained set of forty reusable prompt variants, refilled in the weekly review slot. Where: 5.2, 12.5.

Variant — One prompt-and-settings combination inside a batch, tracked so its survival can be measured. Where: 5.2, 5.4.

Batch — One production run, sized from the survival rate rather than from ambition. Where: 5.4, 5.6.

QA survival rate — Frames passing QA divided by frames generated; the input to batch sizing. Where: 5.6.

QA rejection rate — One minus the survival rate; the diagnostic that says a pipeline layer regressed rather than the model. Where: 5.6, 7.5.

Human time — The largest total line in the pipeline cost table, and the line operators omit. Where: 7.5, 0.4.

Distribution, monetization, and metrics

Cost per usable frame — Generation spend divided by frames passing QA; the metric that makes tier changes comparable. Where: 5.6, 7.5.

Cost per published post — Monthly pipeline spend divided by posts published; below roughly $6 is healthy, above $15 means step down. Where: 0.4, 7.5.

ARPU — Average revenue per user per period; in the book's worked month, $4,230 across 60 subscribers, or $70.50. Where: 10.4, 12.5.

LTV — Blended ARPU multiplied by average subscriber lifetime in months; $70.50 across five months gives $352. Where: 10.4.

Churn — The rate at which subscribers leave; one month less of average lifetime moved LTV from $352 to $585. Where: 10.4, 12.5.

PPV — Pay-per-view: one-off paid content sold inside a subscription, priced at about $8 to $40 per item. Where: 9.4, 10.4.

Subscription tier — A recurring access level priced for entry, mid, or high intent. Where: 9.4.

Revenue hierarchy — The earning routes ordered easiest to hardest, with operator hours attached to each. Where: 9.1.

Media kit — The one-page document answering a brand's capacity, compliance, and rate questions before they ask. Where: 9.5.

Brand deal — A paid deliverable for a company, quoted as units with usage rights and exclusivity written in. Where: 9.5, 11.3.

AI UGC — Generated creative produced to a brand's brief and priced per deliverable, needing no audience of your own. Where: 0.3, 9.5.

Affiliate — Promoting someone else's product for a commission: the cheapest first dollar, the weakest position. Where: 9.1, 9.7.

Merchant of record — A platform that sells on your behalf, remitting VAT and issuing the invoice in its own name. Where: 9.4, 11.8.

Ownership channel — A surface you control, such as an email list, so the audience survives a platform decision. Where: 10.1, 12.4.

Law, compliance, and provenance

Provider — Under the AI Act, whoever develops an AI system or places it on the market under its own name. Where: 11.1.

Deployer — Under the AI Act, whoever uses an AI system under its own authority; you, every time you send a prompt. Where: 11.1.

AI Act Article 50 — The transparency article: 50(2) marks output and binds the provider, 50(4) binds you. Where: 11.1, 11.3.

Deepfake — Content resembling existing persons or events that would falsely appear authentic; the definition triggering 50(4). Where: 11.1.

Disclosure template — The five ready fragments placed by one rule: unavoidable, not available. Where: 11.3.

Machine-readable marking — Provider-side marking as C2PA credentials or pixel watermarks; not your obligation and not reliable. Where: 11.1, 11.5.

C2PA — Content Credentials: a signed manifest describing what created a file and what was done to it. Where: 11.5.

Provenance metadata — Generator, prompt, seed, references, timestamp, and checksum recorded per published asset. Where: 11.4, 11.5.

Publicity rights — A person's right to control commercial use of their likeness, separate from copyright. Where: 11.4, 11.7.

Likeness right — The same interest framed as a right in one's own image; in Poland, prawo do wizerunku. Where: 11.7.

Model release — Written permission from a human appearing in a reference, scoped by territory, term, and media. Where: 11.4.

Data subject rights — Access, rectification, erasure, portability, and objection under Articles 15–22. Where: 11.7.

Retention period — The defined lifespan per data class: 12–24 months for logs, five years for invoices. Where: 11.4, 11.7.

DPIA — A data protection impact assessment, required for large-scale profiling such as a chat persona. Where: 11.7.

Consequence-scanning — Reviewing a planned post for who it could harm or mislead before publishing it. [FIELD] Where: 11.6.

Never-publish list — The written set of things the persona will not do, including political deepfakes. Where: 11.6, 12.4.

VAT OSS — The One Stop Shop: one registration and one quarterly return once the €10,000 threshold is crossed. Where: 11.8.

Withdrawal right — The EU consumer's 14-day right to cancel, lost only when consent and confirmation are documented. Where: 11.8.

spółka z ograniczoną odpowiedzialnością — The Polish limited company; the form the book's compliance material assumes. Where: 11.7, 12.4.

Book conventions

Honesty marker — A bracketed claim label: [UNVERIFIED], [VENDOR CLAIM], [UNTESTED], [SNIPPET], or [AS OF 2026-10]. Where: STYLE.md §6.

[UNVERIFIED] — The marker meaning the claim has no corroborating source you can check. Where: STYLE.md §6.

[FIELD] — This appendix's own annotation, not an honesty marker: field vocabulary covered at category level rather than named in a chapter.

How to use this with the rest of the book

Read the section matching your current layer rather than the whole list; the vocabulary is grouped by where the decision lives. When a batch fails and you cannot say which variable moved, come back and name it: the naming is the diagnosis, and nine of the twelve defects in 5.6 are fixed somewhere other than the frame you are looking at.

Two caveats. These definitions are one sentence, and one sentence cannot carry a failure mode — the chapter reference is the real entry. Every term marked [FIELD] is vocabulary from the wider field rather than from the manuscript, so verify those against the tool's own documentation before relying on one.