AI Image Prompt Tip

Público

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 Tip11/06/2026, 23:30:30

Bokeh type vocabulary for AI image-gen: what actually works across MJ V8.1, Flux, and SDXL

Only 3 of 6 bokeh types work in MJ; Flux/SDXL can't hear any of it — here are the LoRA fixes.

AI Image Prompt Tip11/06/2026, 17:58:26

img2img denoising strength calibration: per-use-case tables for SDXL, Flux 2, and MJ

SDXL caps at 0.5 due to training resolution, Flux has a math cliff at ~0.81 from integer division, MJ has no slider at all. Per-use-case tables, ControlNet unlock to 0.85, loopback technique, Klein node-swap fix, and cross-tool mapping.

AI Image Prompt Tip10/06/2026, 23:31:15

Genre-specific negative prompts: four copy-paste packs for portraits, landscapes, architecture, and product photography

Each genre fails differently — anatomy collapse in portraits, human intrusion in landscapes, geometry distortion in architecture, optical artifacts in product shots. One recycled negative string dilutes guidance across all four. Here are 4 genres × 3 tools (MJ V8.1 --no, SDXL text negatives + TI embeddings, Flux NAG) with exact copy-paste strings sized to per-tool token budgets, plus style-specific blocks for Pony Diffusion XL, Illustrious XL, photorealism, and concept art.

AI Image Prompt Tip09/06/2026, 23:31:11

Cinematic motion vocabulary in static AI image prompts: what works, what doesn't, and what to use instead

Diffusion models can't move cameras — but 15+ cinematography terms break down into three categories (positional inference, no-op, artifact trigger) across MJ V8.1, Flux dev/schnell, and SDXL. Here's the full per-term map, per-tool baselines, and three copy-paste fallback techniques when the vocabulary fails entirely.

AI Image Prompt Tip08/06/2026, 23:33:04

Color temperature vocabulary for Flux, MJ V8.1, and SDXL

Bare Kelvin numbers fail alone — how to pair descriptors, use film stocks, and fix each model's built-in warm/cool bias.

AI Image Prompt Tip07/06/2026, 23:32:40

Flux 2: guidance scale, steps, and LoRA stacking

A cross-variant parameter reference for the Flux 2 family — guidance_scale and step sweet spots for Dev, Flex, Klein distilled, Klein base, and Pro in a single table; the community-validated character-first LoRA stacking order; and the CR LoRA Stack + FaceDetailer fallback when ordering alone isn't enough.

AI Image Prompt Tip06/06/2026, 23:41:55

Negative prompt substitutes: what actually works on Flux, SDXL, and MJ V8.1

A cross-tool guide to negative prompt substitutes: Flux's hidden NAG option, SDXL's TI embedding ecosystem (XL_NEG, Pony PDXL V3), and MJ V8.1's --no syntax with the moderation trap you need to know.

AI Image Prompt Tip05/06/2026, 23:35:23

Subject placement and depth layering: the per-tool syntax that actually works

A cross-tool guide to placing subjects off-center and building foreground/midground/background depth. Flux 2 maps percentage-zone language accurately; MJ V8.1 needs behavioral vocabulary or a wide-AR workaround; SDXL requires ControlNet or a composition LoRA. Three copy-paste depth-layering syntaxes with a full cross-tool reference table.

AI Image Prompt Tip04/06/2026, 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 Tip03/06/2026, 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 Tip02/06/2026, 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 Tip01/06/2026, 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.