Featured case 03 · Deep learning and desktop application engineering
A trainable image model with honest quality evidence
A real caption-conditioned image-model training and editing workflow with owned synthetic fixtures, held-out evaluation, checkpoint safety, and private local workspaces.
Problem
A customizable image tool should prove that training, inference, editing, and dataset governance work together. It must also distinguish a trained demonstration model from a useful general-purpose image generator.
Constraints
Training data needs provenance and permission. Unsafe model serialization must not execute downloaded code. Evaluation data must remain separate from training. Account imports and outside providers must be deliberate.
Design
Python implements caption embeddings, a compact diffusion UNet, EMA, evaluation, and safetensors checkpoint/resume. A TypeScript interface exposes local generation, image-to-image, inpainting, and expansion; known pixels are preserved. Dataset records track rights, deduplication, splits, and AI-origin quarantine. Local accounts protect locked workspaces; an opt-in account connector is separate from native training.
Verification
Tests cover training/inference mechanics, malformed checkpoints, dataset boundaries, image edits, account isolation, and actual bundled services. Provider tests do not spend credits or import arbitrary remote artwork. See private validation and the repository’s sample evidence.
Measured result
The reference custom model completed 6,000 CPU optimizer steps on owned procedural shapes with held-out loss and saved checkpoints. Rough generated samples demonstrate the trained pipeline. This is model-execution evidence, not a comparative quality benchmark.
Limitations
Small synthetic training is not photorealistic generation or broad artistic capability. Imported Stable Diffusion/LoRA support is limited to documented architectures. No third-party training artwork is distributed. CPU desktop packages are unsigned; unlocked content trusts the host OS. Live provider quality and secure public hosting are not claimed.