Style Lab helps you discover stable, repeatable looks by running controlled experiments and comparing outputs side‑by‑side. It focuses on anime and comic styles but works for any visual direction.

Core capabilities:

• A/B and grid comparisons across prompts, LoRAs, samplers, and seeds
• Two‑axis style matrices (e.g., LoRA weight vs. CFG; prompt token vs. sampler)
• Reproducible presets with fixed seeds and shared metadata
• Packaging of winning combinations into LoRA packs

1. Choose a base model

• Anime: a toon/anime‑trained checkpoint
• Comics: a line/ink or stylized checkpoint

2. Lock constants (for reproducibility)

• Seed: fixed (e.g., 123456)
• Resolution: 768–960 short side (anime ~768; comics ~896)
• Negative prompt: text, watermark, signature, low quality, jpeg artifacts, blurry, extra limbs
• Steps: 20–32
• Sampler: DPM++ 2M Karras (good default)
• CFG: 4.5–6.5 (start at 5.5)

3. Define two variables to sweep (2×3 or 3×3 grid)

• Example A: LoRA weight [0.4, 0.6, 0.8] × prompt token presence [none, “cel‑shade”, “manga screentone”]
• Example B: Sampler [DPM++ 2M, Euler a, UniPC] × CFG [4, 6, 8]

4. Generate the grid and compare

• Pick winners based on line quality, color harmony, character fidelity, artifact rate

5. Save a preset

• Record prompt, negative, seed, model hash, sampler, steps, CFG, LoRA(s) and weights
• Tag with purpose: “anime-clean-lines” or “comic-halftone-high-contrast”

• Prompts: add/remove 1–2 style tokens at a time; avoid stacking >4 style tokens
• LoRA strength: style LoRAs 0.4–0.9; character LoRAs 0.2–0.5; stack max 2–3 LoRAs
• Sampler: DPM++ 2M Karras (clean), Euler a (sketchy/ink), UniPC (balanced)
• Steps: 18–28 for fast iterations; 28–40 for detail (diminishing returns >40)
• CFG scale: anime 4–6; comics 5–8 (watch over‑saturation >8)
• Img2img denoise: 0.2–0.4 for refinement; 0.45–0.6 for stylistic shift
• ControlNet/Adapter: lineart or canny 0.5–0.8 weight; start 0–0.1, end 0.8–1.0
• Aspect/size: square for baselines; test AR in a separate sweep; high‑res pass 1.5–2× at denoise 0.2–0.35

Goal: clean linework, flat shading, saturated palettes.

Baseline prompt

• masterful anime key visual, clean lineart, cel shading, vibrant colors, dramatic lighting
• Negative: watercolor, painterly, crosshatch, sketch, text, watermark

Settings

• Sampler: DPM++ 2M Karras; Steps: 24–32; CFG: 5–6
• LoRA: anime‑style pack 0.5–0.8
• Optional ControlNet lineart (preprocessor: lineart anime) weight 0.6–0.8

Grids to try

• LoRA weight [0.5, 0.65, 0.8] × token [no “cel‑shade”, “cel‑shade”, “cel‑shade, colorful shadows”]
• Sampler [DPM++ 2M, Euler a, UniPC] × CFG [4.5, 5.5, 6.5]

Acceptance checks

• Line consistency around eyes/hair tips
• Edge cleanliness after x2 upscale (denoise ≤0.3)
• Saturation holds without banding

Goal: strong inks, halftones, dramatic contrast.

Baseline prompt

• high contrast comic panel, bold inking, crosshatching, halftone shading, screen tones, dramatic pose
• Negative: soft focus, watercolor, oil paint, photorealistic skin pores

Settings

• Sampler: Euler a (for expressive lines) or DPM++ 2M (cleaner)
• CFG: 6–8; Steps: 24–36
• Optional: ControlNet lineart or canny 0.5–0.7; texture LoRA 0.3–0.6

Grids to try

• Halftone token [off, subtle, heavy] × CFG [5, 7, 8]
• LoRA [inking 0.4, 0.6, 0.8] × sampler [Euler a, DPM++ 2M, UniPC]

Acceptance checks

• Legible hatching without moiré
• No posterization on gradients
• Speech bubble/text avoidance (keep negative: text, caption, watermark)

• Fix seed, resolution, and negatives before any sweep.
• Only change one axis per concept (e.g., sampler vs. CFG). Avoid mixing sampler and steps in the same axis.
• Prefer 2×3 or 3×3 grids for quick reads; expand only after a winner emerges.
• Rank images on 3 criteria: fidelity to brief, artifact rate, repeatability.
• Keep a small validation set (2–3 different prompts) to confirm the winner generalizes.

• Preset name: domain-purpose-version (e.g., anime-clean-lines-v3)
• Save: prompt, negative, seed, base model + hash, sampler, steps, CFG, LoRA names + weights, ControlNet settings, upscaler + denoise
• Screenshot the winning grid; export a JSON/YAML sidecar for reproducibility
• Promote stable presets into a LoRA pack to share with your team

• Muddy lines: lower denoise; switch to DPM++ 2M; reduce LoRA stack or weight
• Overbaked style (everything looks the same): drop style token count; reduce CFG; mix a lighter LoRA
• Color banding: increase steps slightly; enable high‑res pass with denoise 0.2–0.25
• Hand/face artifacts: add targeted negatives; use a character LoRA at low weight (0.2–0.4); try a refiner pass
• Halftone moiré: change scale/angle tokens or downscale slightly before final upscale

Once a style is proven across prompts and seeds, package its components into a LoRA pack:

• Include: LoRA files + recommended weights, base model notes, prompt snippets, negatives, sampler/steps/CFG, ControlNet presets, upscaling guidance
• Provide quick‑start grids and a validation prompt list
• Version the pack when updating any dependency

Sharing LoRA packs ensures teams can reproduce the exact anime or comic look without re‑running the entire Style Lab.