FML · Internal R&D · 2026 · Confidential

Rendering
animation
with AI.

ProjectMozzies — internal IP
MethodPrevis-driven AI video generation
ApplicationTom & Jerry — Steer Studios
Scroll

01 / 05

Mozzies —
the experiment

An internal FML project — models and animation locked, but never rendered. We used it to test whether AI could replace the back half of a 3D pipeline entirely.

What went in

DesignFoliage concept
ModelFoliage model
LookdevTree bark
LookdevCattails
Layout · Art direction in progressLayout drawover
LookdevRock asset
LookdevPlants
LookdevTree with moss
LookdevPlant assembly
RenderFinal render
RenderRender test
Render · Comp variations Comp WIP

Character models

Both beetles were fully modelled and rigged before any AI was involved. Their identity is what transferred most faithfully through the pipeline.

Blue Beetle 3D Turntable
Pink Beetle 3D Turntable
Characters — held well

The beetles stayed true

Character identity transferred reliably. Proportions, colour, and surface quality remained recognisable shot to shot when strong identity sheets were used as references.

Environment — loosely interpreted

The world drifted

Despite the depth of the environment build, the tools generated their own interpretation. The vibe holds. The specifics don't. This is the core trade-off.

PREVIS → OUTPUT

Greyscale animatic in · Styled render out

Previs Input 3D Animatic

Greyscale geometry · Final camera moves · No materials · No lighting

AI Render Output Higgsfield / Seedance 2.0

Pixar-style · AI materials & lighting · Shot-by-shot generation

02 / 05

The pipeline
difference

AI replaces the back half. Everything up to animation stays the same.

Standard pipeline

01Concept & DesignShared
02ModellingShared
03Rigging & AnimationShared
04Texturing & ShadingStandard
05LightingStandard
06RenderingStandard
07Compositing & GradingStandard

AI pipeline

01Concept & DesignShared
02ModellingShared
03Rigging & AnimationShared
04Texturing & ShadingReplaced
05LightingReplaced
06RenderingReplaced
04Hero Still GenerationAI — New
05AI Video GenerationAI — Core
06Compositing & GradingHybrid

03 / 05

How it
ran

Shot-by-shot. Every generation is a creative decision — there is no automated pipeline.

01

Split the previs into individual shots

Long takes lose consistency mid-generation. 4–6 second shots hold quality. Each treated independently.

Shot length

4–6 sec

02

Generate hero stills first

Lock the look cheaply before committing video credits. Approved stills become start/end frame anchors for the video.

Model

Nano Banana 2

03

Process the animatic — motion only

Raw previs contaminates the output. Desaturate, blur, or B&W it first. Style comes from the hero still.

Method

B&W / blur

04

Generate: start + end + motion reference

Both ends locked, animatic drives the camera path. Character sheets included as additional references throughout.

Primary model

Seedance 2.0

05

Review, iterate, regenerate

3–8 attempts per shot. Each failure informs the next — model choice, reference processing, prompt language.

Avg attempts

3–8 per shot

04 / 05

Findings &
limitations

Earned from hours of generation — not theoretical. Real capability and real limits in equal measure.

Worked

Hero still → video anchor

The single most effective quality control. Lock the look first, iterate cheap, then commit.

Worked

Start + end frame locking

Both ends anchored dramatically reduces mid-shot drift.

Worked

Processed animatic

B&W or blurred previs gives motion data without contaminating the aesthetic.

Worked

Short shots

4–6 second clips held quality far better than longer continuous sequences.

Failed

Seedance as cinematic

It's an e-commerce model. Kling 3.0 and Cinema Studio performed better for character shots.

Failed

Raw animatic as reference

Model reads geometry as a style target. Always process first.

Learned

Content filters

Stylised cartoon content trips IP detection regularly. Minor image processing usually resolves it.

Learned

Model selection per shot

Environment reveals, character close-ups, and action sequences each favour different models.

Learned

Creative direction, not automation

Every decision is a creative call. Experience and taste directly affect output quality.

Hard limits

Character consistency across shots

Identity drifts without a trained character model. Careful reference management required — not fully solvable today.

No frame-accurate camera matching

The model interprets the animatic, it doesn't parse it. Exact timing and easing will deviate.

Non-deterministic output

Identical requests produce different results. Once a take is approved, treat it as locked.

Credit cost at scale

Generation costs are real and should be budgeted explicitly — though still well below a full re-render.