Wang Jianjun
中文
All projects

Multimodal AI Production System

Bring commercial-image generation, local editing and quality checks into one workflow.

I designed the AIGC workflow; per-image cost was projected to fall from ¥89.8 to ¥43.9, with actual outcomes still unverified.

View diagram (swipe sideways)
Multimodal AI Production System cover diagram
AI Product Manager / AI Application EngineerMay – Aug 2026
MultimodalAIGCAdobe UXP+4

Key results

  • Projection: 10 → 30 images/day
  • Projection: ¥89.8 → ¥43.9 per image (−51%)

My role and contributions

At the end of my participation, all output and cost figures were still project projections. Pilot recommendations are retrospective; subsequent production outcomes have not been verified.

As AI product manager / AI application engineer, I designed the workflow across the Photoshop plugin, backend API, DAG scheduling, generation and quality evaluation.

30/day

Projected output

+200%

¥43.9

Projected per-image cost

−51%

3

Pipeline stages

2

Quality loop steps

Challenge

Generation alone does not solve commercial consistency, editability or the delivery workflow.

Solution

I decomposed the problem into Understand → Draft → Edit. Multimodal understanding captures brand constraints; generation produces drafts; editing returns to a controlled commercial workflow.

1

Understand

LookSpec captures brand constraints structurally so generation has something to ground on.

2

Draft

Generation produces candidate drafts quickly, shortening the trial-and-error cycle.

3

Edit

BBox-driven local edits bring output back into a controlled commercial delivery loop.

System path

The Photoshop plugin submits work through the backend API. DAG scheduling runs generation and quality checks before designers edit and review the result.

Technical details

  • Adobe UXP plugin for in-Photoshop editing handoff.
  • LookSpec structured understanding of brand constraints.
  • BBox localization for precise local edits.
  • Automated quality checking with human review in the loop.

Results

Projected daily output

10/day30/day

Projected per-image cost

¥89.8¥43.9

Projection basis

When my participation ended, a category pilot was still needed to establish actual output, human edit rate and cost baselines. This case study does not document subsequent production verification.

Why it matters

The workflow links brand constraints, editable output and quality review; the projected cost advantage still needs a real pilot.

Lessons learned

The retrospective recommendation is to run a focused category pilot and verify human edit rate and business adoption before scaling similar projects.