AF7620 Chap.7 Style and Character Consistency in AI Film
Style and Character Consistency in AI Film
The problem the reading sets out to solve
The reading assigned in the technical pipeline lecture is a short published study presented at a leading computer graphics conference in 2025. It begins from a problem every student meets when making an AI film: text-to-image diffusion models make visual experiment easy but often fail to keep style and characters consistent across scenes, which leaves a gap between the filmmaker's intent and what the model can deliver.
The study compares workflows developed while making two fully AI-generated short films and proposes a framework oriented to filmmakers.
Three pipelines, measured
The study measured three generative pipelines. Prompting alone gave medium style consistency, low character consistency and an average of about twenty-one per cent stable frames.
A single LoRA, a lightweight fine-tuned adapter, gave high style consistency, medium to low character consistency and about forty-six per cent stable frames. Several LoRAs used together gave high style consistency, medium to high character consistency and about seventy-one per cent stable frames.
The main limitation of the multi-adapter approach is practical: it needs technical skill and a capable graphics processor.
First film: style held by self-reference
The first film recreated the look of Song Dynasty ink painting.
Its workflow used a commercial image generator with a style-reference feature, feeding each generated image back as the visual prompt for the next so that the look stayed coherent and abrupt shifts were avoided. Prompts were controlled to avoid a synthetic look, and because the painting tradition values spiritual resemblance over likeness, character identity was treated loosely.
Prompts described clothing, age and other visible traits separately for wide, medium and close shots to reduce drift. Even so, the study reports that narrative coherence came more from editing and framing than from the model.
Second film: style and identity separated
The second film aimed for both a consistent style and a character who stayed recognisably the same.
A single adapter trained on a mixed set of style and character images failed, because it could not separate the two, and the character's appearance was unstable. The team instead trained three adapters: one for real-world backgrounds, one for dream sequences and one for the character.
They trained the style adapters first, generated a style-consistent seed image of the character, expanded it by hand into a set of poses and angles, and trained the character adapter on that set.
The result was more stable, identity-preserving output across shots, and because each adapter is parameterised independently, characters and backgrounds can be adjusted separately at any shot distance.
What this means for your pitch
The proof-of-concept asset and the pitch deck both show characters more than once.
The study gives you a defensible reason to budget time for training a character adapter, and a measured result to cite when you explain why.
What this chapter covers
- 01
Why diffusion models drift between shots
- 02
Stable frames under three pipelines
- 03
Style reference and self-referencing
- 04
Prompting by shot distance
- 05
Why a mixed-data adapter failed
- 06
Separate adapters for backgrounds, dreams and the character
- 07
Costs of the multi-adapter route
- 08
Using the result in a pitch
Worked example · free
Choose a consistency pipeline for a proof-of-concept
- 3Compare the three pipelines against the constraint.
- 3Choose and justify one route.
- 2State the limit honestly for the pitch.
Key terms
- Style Reference
- A feature of some image generators that uses an existing image as a visual prompt, so that new images inherit its look.
- Stable Frame
- In the study's measure, a generated frame in which style and character remain consistent with the rest of the sequence.
- Character LoRA
- A fine-tuned adapter trained on many images of one character so that the character's identity is reproduced reliably across shots.
- Style LoRA
- A fine-tuned adapter trained on images in one visual style, used to keep backgrounds and scenes looking consistent.
- Seed Image
- A starting image from which further variations are generated or a training set is expanded, used to keep a series anchored to one look.
Style and Character Consistency in AI Film FAQ
What is the difference between a style adapter and a character adapter?
A style adapter learns how things look, such as a watercolour texture or a dream-like palette, and applies that to any subject. A character adapter learns who one character is, their face, build and costume, so that the same person appears in any scene. The study found that training them separately worked where one mixed adapter failed.
Is prompting alone ever enough?
For one-off images or a style that tolerates loose likeness, it can be. The study's first film used prompting with a style reference and accepted loose character identity because its painting tradition allowed it. For a narrative with a recognisable lead, prompting alone kept only about a fifth of frames stable, which is why the second film moved to adapters.
Assessment move
Run the study's comparison on your own character: generate the same character ten times with a prompt only, then with a reference image, and count how many results you would accept. Recording your own stable-frame share gives the pitch a measured basis for the pipeline you propose.
Working through Style and Character Consistency in AI Film in AF7620? Sia is AskSia’s AI Arts and Humanities tutor — ask any AF7620 Style and Character Consistency in AI Film question and get a clear, step-by-step explanation grounded in how AF7620 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
