Character Concept · Semi-realistic · AI Lab

Pinkdot Lolita — Character Concept

Stack

Role: Concept direction, generation, per-zone detail regeneration & composite
Base model: Krea 2 (open-weights)
LoRA: Self-trained stylized-realism style LoRA ↓
Tools: ComfyUI, self-trained style LoRA on Krea 2, directed generation; splash generated with a self-trained splash LoRA on Krea 2, Qwen-Image and GPT Image zone edits; per-zone detail regeneration; Photoshop composite; idle loop rigged from layer-split parts (Qwen-Image clean plates, custom displacement rig)
Year: 2026

Process

Pink polka-dot lolita set — bonnet, corset dress, lace socks and platform Mary Janes — shown in a splash illustration, its looping idle animation, and three poses. Hands, lace, shoes and ribbons were finished zone by zone.

From the AI Lab

Tool Look Detailer — canvas-to-ComfyUI repair skill One command from a Photoshop lasso selection to a regenerated, pixel-aligned detail layer — a scripted bridge between the live canvas and ComfyUI.

Problem it solves

Generated character art ships with local defects: chainmail painted as dead-pixel mush, melted fingers, buckles that read as blobs. Regenerating the whole image to fix one patch is a lottery — you lose the parts that were already right. Manual paintover works but is slow at scale. The fix should be surgical: regenerate only the broken region, at higher resolution than the canvas itself, and keep everything else untouched.

What it is

A scripted skill that runs against the live Photoshop session. The workflow from the artist’s side is two gestures: draw a rectangular lasso around the defect, invoke one command with a short brief (“chainmail”, “fix the hand”). Everything else is automated:

Look Detailer — a square selection on the gold brocade obi, live on the canvas

Capture. The script reads the selection bounds from the open document via Photoshop scripting and saves a flattened composite — the original layers are never touched.
Crop & route. The selected region is cropped, aspect-matched, and dispatched to a self-built detailer workflow on a local ComfyUI server — three interchangeable variants tuned for different repair jobs, picked by command suffix.
Defect-focused prompting. Each run gets a written brief in a fixed grammar: name the defect, state the develop action, list what must be preserved, lock the painterly style. The model regenerates the patch supersampled at 2K.
Pixel-aligned return. The result is resized back to the exact selection size and placed as a new smart-object layer at the selection’s position, corrected for DPI and rounding so it lands within a pixel of the source region.

The run from the selection above — a 380 px square around the obi, regenerated as a 1024² patch at denoise 0.30 with the same style LoRA the piece was made with. The brocade motifs and thread texture sharpen and the braided cord’s strands separate, while shape, colour and position stay put (left: selection, right: output):

The run's actual output — the obi selection before and after, 1024² each

Guardrails that matter

Non-destructive by contract. Results only ever arrive as new layers; the original pixels stay intact underneath.
Full patch, no auto-mask. The tool never feathers or hides its own seams — blending is an art decision, so masking stays in the artist’s hands.
One invocation = one generation. No hidden multi-seed retries; a retry is an explicit, separate decision.
Honest QA before placement. Every result is inspected at full resolution — zoom crops, finger counting on hands, style-drift check against the source brushwork. A defective result gets rejected and re-briefed, not silently placed.

Used on

Same-session repairs on this site’s posted works: the chainmail gaps and warhammer on the Knights avatar set and the belt rig and leg plate on its female knight — each one a lasso, a one-line brief, and a placed layer.

Open in LAB →