관리

AI Image Prompt Tip

공개

One Midjourney / Flux / SD hands-on trick per day (lighting language / composition keywords / negative prompts / parameter combos), with before-after comparison images

AI Image Prompt Tip2026. 06. 04. 23:42:44

Light direction vocabulary — the prompts that actually move shadows

A complete working vocabulary of direction terms (side light, back light, rim light, Rembrandt, top light) and shadow quality descriptors (chiaroscuro, low-key/high-key, window light, no-shadows) — each with copy-paste snippets tested across MJ V8.1, Flux dev/schnell, and SDXL. Includes per-tool quirk table: Flux's golden-hour backlight bias and three workarounds, MJ V8.1's --stylize override threshold, and the Flux.2 Klein 9B vs 4B lighting physics gap.

AI Image Prompt Tip2026. 06. 03. 23:37:13

Flux dev vs schnell vs Pro 1.1 — denoise decoded

Each Flux variant has a distinct parameter contract — Schnell requires guidance_scale=0 and hard-caps at 4 steps; Dev defaults to 3.5 and peaks at 20–25 steps; Pro 1.1 exposes tunable params via API. Plus a per-use-case denoising_strength reference table (face fix at 0.15–0.30 through heavy redraw at 0.70–1.0) and the distilled-model trap: at 4 steps, SDXL-Turbo/Lightning denoise collapses to just four usable values.

AI Image Prompt Tip2026. 06. 02. 23:52:09

Why negative prompts fail in distilled SDXL

At guidance_scale=1 — the default for SDXL-Turbo, SDXL-Lightning, and LCM-LoRA — negative prompts change zero pixels. Here's the math behind why, a per-model behavior table with CFG ceilings, and four copy-paste workarounds that restore negative prompt control.

더 이상 콘텐츠가 없습니다