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 Tip2026/05/29 23:31:52

Hidden knobs: VAE, guidance_scale, and clip_skip

VAE, guidance_scale, and clip_skip are the three parameters most users leave at defaults — each one has a correct value per model family, and getting any wrong causes specific, diagnosable artifacts. This tip covers sdxl-vae-fp16-fix as the only safe SDXL VAE for fp16 inference, Flux guidance_scale split by subject type, and clip_skip behavior across SD 1.5 / SDXL / SD3 / Flux — with a unified cheat sheet.

AI Image Prompt Tip2026/05/28 23:53:34

Negative prompts are dead — here’s what actually works in SD3, Flux, and MJ V8.1

SD3/SD3.5 never learned to use negative prompts — passing one injects noise, not filtering. Flux’s entire API family (7 endpoints audited) has no negative_prompt field at all. MJ V8.1 supports --no but multi-prompt :: syntax is unsupported, making ::-1 unavailable. Per-tool replacements: positive reframing + SLG for SD3, guidance_scale tuning for Flux, and tight --no lists for Midjourney.

AI Image Prompt Tip2026/05/28 00:00:29

Aspect ratio and canvas shape: a cross-tool syntax guide

How to set aspect ratio and canvas shape in Midjourney V7/V8.1 (--ar), Flux dev/schnell/Pro (width/height API params), SDXL (resolution buckets + micro-conditioning), and SD3/SD3.5 — with a cross-tool copy-paste reference table and use-case ratio lookup.

これ以上のコンテンツはありません