Selected work

Public reference · Generative AI

Databricks LoRA Creative Asset Pipeline

A GPU-oriented workflow for adapting SDXL to an existing visual language, generating traceable variations, filtering failures, and handing candidates to designers.

Type
Creative AI production workflow
Role
Architect / engineer
Status
Public reference
Focus
Model adaptation, traceability, human review
SDXLLoRADiffusersPEFTBLIPMLflowDelta LakeVision scoring
System map

The operating path.

A deliberately simplified architecture view. The case study below explains where the important guarantees and decisions live.

  1. 01Reference assets
  2. 02BLIP captions
  3. 03LoRA training
  4. 04MLflow + prompts
  5. 05Batch generation
  6. 06Quality + design handoff

The problem

Why this system needed to exist.

Design teams often need many controlled variations of a familiar asset type. Generic image generation can drift away from the established visual language, lose prompt and seed provenance, and create more review work than it removes.

The useful system is not only a fine-tuned model. It needs repeatable training, traceable generation, quality filtering, ranking, and a handoff designers can actually use.

The system

How the pieces work together.

Reference assets are captioned with BLIP and prepared for SDXL LoRA adaptation using Diffusers, PEFT, and Accelerate. MLflow tracks parameters, metrics, artifacts, and adapter weights.

A Delta-backed prompt library drives seeded batch generation. Vision scoring filters obvious failures, human review controls approval, and the handoff exports selected assets with manifests, contact sheets, provenance, and line-art previews. The original workflow helped designers accelerate repetitive exploration while preserving final human judgment.

01

Reproducible generations

Prompt dimensions, seed, LoRA scale, model run, and review state remain attached to each candidate rather than disappearing after an image is created.

02

Quality gates before handoff

Automated scoring removes obvious failures and prioritizes candidates, while designers remain responsible for the final creative decision.

03

A deliverable, not a notebook cell

Approved files arrive with manifests, contact sheets, metadata, and previews that support real design work and future reproduction.

Source boundary

The public repository replaces original game assets and brand details with synthetic reward-chest examples. It documents which pieces are source-derived and avoids inventing generation-quality, runtime, cost, or quantified design-savings claims.

A system like this on your roadmap?

Let’s make the hard parts explicit.

Discuss the project