Continuity

Week 8 · SLOP8760

Phase 3 of 4 · Holding It Together

This week adds

Continuity as a system: a reference library built once, a model sheet that documents it, and an audit that checks every new shot against both.

Two ways to lock identity

10,000xfewer trainable parameters than GPT-3 175B fine-tuned with Adam

Reference adapters need no training and hold loosely. LoRA trains first and holds harder: it freezes the pretrained weights and injects trainable rank decomposition matrices into each layer instead. The ratio is that paper’s own language-model measurement, not a figure for a diffusion checkpoint.

Source: LoRA: Low-Rank Adaptation of Large Language Models — https://arxiv.org/abs/2106.09685

There are two genuinely different ways to make a generator produce the same face repeatedly. Reference adapters — the kind Phase 2 already used to carry identity from an image into a new generation — need no training and work the moment a reference image exists, but they hold identity loosely enough that drift is still possible under enough motion or style pressure. LoRA and DreamBooth work the opposite way, and they are not two names for one method. Both want a training pass on a few images of the subject before the first shot, and both hold a subject harder than any single reference image does. Where they part is what they train: DreamBooth’s method is to “fine-tune all layers of the model”, so the artefact is a whole checkpoint at a whole checkpoint’s cost, while LoRA leaves the pretrained weights frozen and the artefact is a small file that loads on top of a checkpoint you already have.

The comparison on the slide is the paper’s own, against a full fine-tune of GPT-3 175B: 10,000 times fewer trainable parameters and three times less GPU memory. The mechanism is a rank decomposition, and the paper reports that for the query and value projections “a rank as small as one suffices” — the adaptation a subject needs is much smaller than the model it adapts. For comparison, an IP-Adapter is 22M parameters and needs no training pass at all.

Given only a few (typically 3-5) casually captured images of a specific subject
Ruiz et al., DreamBooth, 2022

Source: DreamBooth — https://arxiv.org/abs/2208.12242

The reference set is the whole system

The top tier hands the model five references: the face, the full body, two dry states, and the street itself as a subject of its own. H3’s reference task takes up to nine images, three clips, and twelve files in total.

Source: MiniMax-H3 model card, Tasks — https://huggingface.co/MiniMaxAI/MiniMax-H3

A model sheet is the document a reference library becomes once it’s organised for someone else to use: a face reference, a full costume reference, and a short list of props or details a viewer would notice if they changed. Building the sheet once, before the season’s first real shot, is what turns “I remember what the costume looks like” into something a second person — or a second week of production — can check against without asking.

The fifth reference on this tier is the one students skip. A scene that runs several shots in the same location needs a master plate of the location itself, locked the same way a face is locked, or a background detail drifts between two shots that are supposed to be seconds apart in the same room. In the recorded prompt it is written as a subject in its own right, with its own retention line, alongside the character.

The card’s limits are worth knowing before you plan a shot around them: up to nine reference images, three reference videos and three standalone audio clips, each 2 to 15 seconds, with a total of no more than twelve files.

The model sheet, then the audit

Face, costume, props, location — a verdict per item, not one verdict for the shot. A shot can pass on the face and fail on a prop in the background at the same time.

Recorded continuity ladder, tier t1: 1 reference: face — Ladder t1 · 1 reference: face

The ladder

One two-shot scene, with more reference behind it each time.

A single face crop, and nothing else
Plus one full-length reference: two in total
Plus the same character dry, half and full: four in total
Plus a scene master plate for the location the two shots share

Before class

Gather whatever reference images already exist for one character — even just a single face crop — to use as the starting point for this week’s ladder.

This week’s exercise

Build all four ladder tiers of reference material, then generate the same two-shot scene against each tier and audit both shots item by item. Bring the four reference sets, the resulting shots, and the audit notes to Wednesday’s Dailies.

Reading

LoRA: Low-Rank Adaptation of Large Language Models
Hu et al.2021
freezes the pre-trained model weights and injects trainable rank decomposition matrices into each layer
DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation
Ruiz et al.2022
Given only a few (typically 3-5) casually captured images of a specific subject
VACE: All-in-One Video Creation and Editing
Jiang et al.2025
reference-to-video generation, video-to-video editing, and masked video-to-video editing
Phantom: Subject-consistent video generation via cross-modal alignment
Liu et al.2025
extracts subject elements from reference images and generates subject-consistent videos following textual instructions
MiniMax-H3 model card
MiniMaxAI2026
Up to 9 reference images, 3 reference videos

Continuity#

Week 8 · SLOP8760#

This week adds#

Two ways to lock identity#

The reference set is the whole system#

The model sheet, then the audit#

The ladder#

Before class#

This week’s exercise#

Reading#