Visual Workflow

AZIT LORA PIPELINE · END-TO-END VISUAL TRACE

Six stages from anchor identity synthesis through conditioned inference, captured live from the working environments (ComfyUI, AI-Toolkit). Followed by a checkpoint comparison gallery showing how the LoRA renders the same prompt across five training checkpoints (Steps 2200 - 3000).

Anchor identity synthesis

Generate the initial subject face that establishes the identity seed.

Bare-face ID canonicalization

Refine stage 01 into a bare-face, passport-style portrait so the identity seed is locked unambiguously.

Multi-view dataset synthesis · i

Feed stage 02 into a ComfyUI workflow to synthesize a multi-angle dataset — so the LoRA learns the full head as a 3D volume, not just the frontal face.

Multi-view dataset synthesis · ii

Continuation of stage 03 — augmenting the dataset with additional angles, distances, and expressions.

LoRA fine-tuning · AI-Toolkit

Feed the stage 03 - 04 dataset into AI-Toolkit to train the LoRA.

Conditioned inference · editorial styling

Load the stage 05 LoRA into a ComfyUI workflow to generate K-Pop editorial imagery.

SUPPLEMENTARY

Additional captures from the working pipeline.

Supplementary capture · pipeline state

CHECKPOINT COMPARISON GALLERY

Steps 2200 - 3000
across 5 checkpoints

Each strip shows the same prompt rendered with five LoRA checkpoints saved during training. Reading left to right: Step 2200 (optimum, best identity-vs-generalization balance) up through Step 3000 (overfit territory). The visual progression confirms what the ArcFace curve already showed numerically — identity locks in by Step 2200; later checkpoints memorize dataset artifacts at the cost of generalization.

Numerical detail in RESEARCH ↗ (03 ArcFace Benchmark Results).