管理

AI Image Prompt Tip

公開

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 Tip2026/06/03 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.

AI Image Prompt Tip2026/06/01 23:36:32

Depth cues that actually work — and the ones AI ignores

Camera aperture values produce zero change in AI image depth of field — natural-language vocabulary like 'foreground bokeh', 'volumetric fog', and 'shallow depth of field' is what actually moves the image. Covers per-tool syntax for MJ V7/V8.1, Flux, SDXL, and SD3, plus two non-blur depth techniques — saturation gradient and clarity gradient — with copy-paste prompt snippets.

これ以上のコンテンツはありません