AI-Assisted
Workflows in
Animation Production

A controlled workflow study that asks where AI should accelerate beginner animation production, where humans should intervene, and how a Live2D pipeline can stay reliable after automation enters the process.

Study Controlled A/B experiment
Sample 20 animation students
Task Slice, clean, rig, test one Live2D character
My focus Workflow design, research synthesis, evidence storytelling

AI made the early stage faster. It also made the pipeline easier to mistrust.

Manual slow / precise

High control, lower hidden-error risk, heavy repetitive labor.

AI-assisted fast / unstable

Quick slicing and setup, but cleanup and checking move downstream.

Design question

How might we structure the AI-human handoff so speed does not erase creative control?

The contrast becomes clearest when the two workflows are seen as production stories.

AI-assisted

Storyboard showing an AI-assisted animation workflow: quick slicing, cleanup work, and quick prototype output.

Manual

Storyboard showing a manual animation workflow: careful slicing, clean rig preparation, and high-quality rigging.

The AI layer was built as an operating tool, not just a one-off model test.

Main workflow overview diagram showing manual and AI-assisted animation production paths.

During the project, I used ComfyUI to assemble an automated layer-separation workflow with the SeeThrough model. The goal was to turn a flat character image into editable layer candidates, then keep the human review step close enough to catch broken edges, missing parts, and naming errors before rigging.

To make the workflow easier for non-technical users, the ComfyUI API was wrapped as a Discord bot. Students could submit artwork from Discord, trigger the AI separation job, and receive exported layer files without opening the node graph directly.

ComfyUI logo.
Platform

ComfyUI workflow

Node-based image processing pipeline for reproducible layer separation experiments.

Model

SeeThrough separation

AI-assisted pass that identifies character parts and produces editable layer candidates.

Discord logo.
Interface

Discord bot access

API wrapper that lets users submit artwork and collect outputs through a familiar chat flow.

01 Upload artwork

Student sends a character image to the Discord bot.

02 Run API job

The bot calls the ComfyUI workflow with fixed model settings.

03 Separate layers

SeeThrough generates hair, face, eye, mouth, and body candidates.

04 Review output

Animator cleans the layers before Live2D rigging.

The workflow maps where AI accelerates production and where human review protects quality.

Human control
Prepare artwork
Inspect layer naming
Adjust rigging
Validate face tracking
AI acceleration
Suggest slicing
Separate parts
Prototype motion
Production output
Clean layers
Live2D model
Motion test

Efficiency

Does a structured AI workflow reduce completion time and perceived effort?

Redistribution

Which stages become lighter, and which stages absorb new correction work?

Quality

Can output stay structurally stable when AI participates in slicing and rigging?

A/B experiment design, rebuilt as a readable evaluation matrix.

Research Question
Evidence
Interpretation
RQ1 / Efficiency and effort
Task time, perceived effort, task logs
AI was expected to lower early-stage production load.
RQ2 / Effort redistribution
Slicing, cleanup, rigging, and testing time
Timing by stage reveals whether AI saves work or moves work.
RQ3 / Quality and behavior
Output review, survey, checking behavior
Reliability depends on whether users verify AI output.

The final story is a trade-off model: speed improved, but supervision became more important.

Quantitative comparison across slicing, cleanup, rigging, testing, total time, survey results, and rating distributions.

Production evidence

The workflow was validated through a Live2D output and face-tracking test, so the case study ends with the production result rather than another research diagram.