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 Tip2026/06/18 23:25:45

SDXL ControlNet: strength + guidance timing

Two parameters, two axes: `control_weight` sets per-step intensity; `guidance_start`/`guidance_end` controls which denoising steps fire at all. Per-preprocessor copy-paste settings for Tile, Depth, OpenPose — ComfyUI and A1111.

AI Image Prompt Tip2026/06/17 23:27:43

Per-LoRA negative management: the type-safety matrix your workflow is missing

Standard negative prompts were designed for blank base models — the moment you load a LoRA, token conflicts and concept bleeding change everything. This article establishes the first systematic per-LoRA negative management framework: a type-safety matrix for style, concept, and character LoRAs; the trigger-word trip wire; concept bleeding thresholds; and per-tool implementation for SDXL (ComfyUI dual-CLIP wiring), Flux dev (NAG, ONG), and MJ V8.1 (--no mechanics and limits).

AI Image Prompt Tip2026/06/16 23:29:35

Focal-length vocabulary for AI image-gen: what works, what doesn't, and what to use instead

Bare mm values are no-ops in MJ V8.1 — three independent A/B tests confirm it. SDXL is the exception. Here's the full per-tool vocabulary map: descriptive terms that work, LoRA workarounds, and copy-paste prompts for telephoto compression, wide distortion, and macro depth across MJ, Flux, SDXL, and SD3.

AI Image Prompt Tip2026/06/15 23:29:40

Two lights, two colors: how to prompt split-tone drama in AI image gen

The three vocabulary tiers (plain color / gel terms / cinema spec) map to different AI training clusters — use the wrong one and two-color setups collapse into single-tone output. Copy-paste prompts for MJ V8.1, Flux, Nano Banana 2, SDXL, and SD3.

AI Image Prompt Tip2026/06/14 23:27:10

Dynamic Thresholding: push SDXL CFG past 8 without color blowout

Dynamic Thresholding (DT) fixes SDXL color blowout at high CFG by clamping latent spread each sampling step — copy-paste settings for ComfyUI and A1111/Forge across four use cases (portrait, landscape, product, extreme CFG), plus a bonus Flux hack.

AI Image Prompt Tip2026/06/13 23:29:33

Style-lock vs content-lock negatives

Negative prompt lists fail when they mix two incompatible jobs — style-lock (blocking rival aesthetics) and content-lock (removing generation defects) require different token types. This article introduces the two-budget framing, maps each to the tokens that actually work, and delivers per-tool copy-paste strings for MJ V8.1, SDXL, Flux dev/Flux 2, and SD3.

AI Image Prompt Tip2026/06/12 23:30:04

Prompt slot hygiene: separating composition and lighting tokens per tool

Each tool's text encoder processes token order differently — here's the slot-by-slot template for MJ, Flux, SDXL, and SD3.

AI Image Prompt Tip2026/06/11 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 Tip2026/06/11 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 Tip2026/06/10 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 Tip2026/06/09 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 Tip2026/06/08 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.