Platform Guide

Each prompt is provided in four platform-specific variants. Understanding their strengths will help you get the best results. The guide also covers three additional platforms — FLUX, Ideogram, and Recraft — worth reaching for when their specialties match your use case; the prompts in this library adapt to them with minor tweaks.

🔵 Nano Banana 2 (Featured — Best Overall Quality)

Nano Banana 2 (powered by Gemini 3.1 Flash Image) is a major upgrade over the original. Treat it as a full-featured creative engine, not a quick-filter tool.

Style: Detailed, structured natural language. Front-load the asset type and primary subject, then layer in style, lighting, and composition details.
Strengths: Combines pro-model quality with flash-model speed. Supports multi-turn conversational refinement, search grounding (reference real-world subjects by name), accurate text rendering, character consistency (up to 5 characters), and output up to 4K resolution.
Why It’s Featured: NB2 consistently produces the most detailed, accurate, and stylistically faithful results across all 72 styles in this library. Its conversational refinement loop means you can iterate toward perfection without re-writing your entire prompt.

Key Tips:

Tip	Details
Be specific, not vague	NB2’s reasoning engine rewards precise descriptors for textures, colors, and lighting over generic adjectives
Front-load important details	Place the asset type and primary subject at the beginning of your prompt
Specify aspect ratios	Explicitly state the format: `16:9` for wallpapers, `3:4` for Instagram, `9:16` for stories, `1:1` for profile pictures
Iterate conversationally	Refine with follow-ups: “Change the palette to warmer tones,” “Move the subject left,” “Add more texture to the background”
Leverage search grounding	For real-world subjects, ask the model to reference accurate imagery: “Use image search to find accurate reference for [specific subject]”
Use positive framing	Instead of “no cars,” describe the scene you do want: “empty, deserted street”
Specify camera/lens	For photographic styles, include lens (85mm, 35mm, macro), aperture (f/1.8, f/8), and camera angle (low, eye-level, bird’s eye)
Name your lighting	Use specific lighting terms: “Rembrandt lighting,” “golden hour,” “tungsten warm glow,” “volumetric fog”

ChatGPT (GPT-4o Image Generation)

Replaces the standalone DALL-E 3 model, which is deprecated as of May 2026.

Style: Fully conversational, natural language. Describe your scene as if briefing an artist.
Strengths: Excellent scene comprehension, precise spatial relationships, and iterative refinement through follow-up messages.
Tips: Be specific about subject placement, lighting direction, and mood. You can assign it a persona (e.g., “Act as a professional product photographer”) for higher fidelity. Use follow-up messages to adjust individual elements without re-describing the whole scene.

Midjourney (V7)

V7 is the current default model. Parameters like --v 6.0 are no longer needed.

Style: Natural language with parameters appended at the end. V7 has moved away from keyword stuffing toward descriptive sentences.

Key Parameters:

Parameter	Purpose	Example
`--ar`	Aspect ratio	`--ar 16:9`, `--ar 4:5`
`--s`	Stylization (0–1000). Low = literal, high = artistic	`--s 250`
`--no`	Exclude elements (replaces negative prompts)	`--no text, blur`
`--cref`	Character reference — maintains identity across images	`--cref [URL]`
`--sref`	Style reference — applies the aesthetic of a reference image	`--sref [URL]`
`--iw`	Image weight for reference influence (0–2)	`--iw 1.5`
`--chaos`	Variation in initial grid (0–100)	`--chaos 30`

Tips: Place the most important visual information first. Use --niji for anime and illustration styles.

Stable Diffusion (SDXL / SD 3.5)

Style: Supports both keyword tags and natural language. Modern models (SDXL, SD 3.5) work best with descriptive sentences, structured as Subject → Action → Environment → Style/Lighting.
Negative Prompts: Use sparingly and only for specific unwanted elements. Bloated negative prompt lists degrade output quality on modern models. Try generating without negatives first, then add targeted exclusions as needed.
Tips: Avoid over-weighting (e.g., (keyword:1.5)) — use it to nudge, not force. Ensure your resolution matches the model’s native aspect ratio.

FLUX (Black Forest Labs)

The FLUX.1 family spans open-weight models ([schnell], [dev]) and hosted pro tiers (FLUX.1 [pro], FLUX 1.1 [pro]), plus the FLUX.1 Kontext suite for image editing.

Style: Descriptive natural language, similar to SDXL/SD 3.5 prompting but with stronger prompt adherence — complex multi-subject scenes and spatial relationships hold together well. No negative prompts on standard pipelines.
Strengths: Excellent prompt following, photorealistic rendering, solid in-image typography, and strong anatomy. Open-weight variants run locally (ComfyUI, Diffusers) and support LoRA fine-tuning.
Kontext (image editing): FLUX.1 Kontext performs in-context generation and editing — prompt with text plus an image to make targeted local edits, preserve a character across scenes, or apply a reference style, all without fine-tuning or masks. Iterative editing keeps the rest of the image stable, making it a strong alternative for the img2img workflows in this library.
Tips: Reuse the Stable Diffusion prompts from each style doc, dropping the negative prompt. For Kontext edits, give short imperative instructions (“Change the background to a rainy street, keep the subject unchanged”) and iterate one change at a time.

Ideogram (3.0)

Best-in-class text-in-image rendering — the go-to platform when your image must contain legible, correctly spelled words.

Style: Natural language. Put the exact text you want in quotes and describe the typography (serif, hand-lettered, neon sign) and placement.
Strengths: Market-leading text accuracy (posters, logos, signage, packaging, book covers — even multi-word headlines), strong graphic-design layouts, and photorealism. Style References (upload up to 3 images) and reusable style codes give precise aesthetic control; Magic Prompt expands terse prompts.
Tips: Ideal for styles in this library that feature typography (concert posters, vinyl album covers, vintage advertising). Pick a design category (Poster, Logo) when available, and spell out the exact wording — e.g., a screen-printed gig poster with the headline "MIDNIGHT CANYON" in hand-drawn psychedelic lettering.

Recraft (V3)

A design-focused model — the only major platform that generates true vector (SVG) output, built for brand and production design work.

Style: Natural language plus explicit design controls — style presets, exact brand colors, and text size/position on the canvas.
Strengths: Native vector art generation (logos, icons, illustrations that scale infinitely), custom brand styles derived from a set of reference images without retraining, accurate text rendering with positioning control, and a full editing suite (vectorize, background removal, upscaling, inpainting/outpainting).
Tips: Reach for Recraft when the deliverable is a logo, icon set, or brand asset rather than a rendered scene. Feed it exact hex/RGB brand colors, and use vector output for anything destined for print or scaling. Styles like Minimalist Notion, Art Deco, and Glass Embossed 3D translate especially well.