Prompts & Prompt System

How Prompts Work

Prompts are text descriptions that guide the AI processor in generating your 360° panoramic image. Better prompts produce better results.

Writing Effective Prompts

Basic Structure

Start with the subject, then add details about style, lighting, and quality:

"a serene mountain lake at sunset, volumetric lighting, highly detailed, 8k"

Negative Prompts

Tell the AI what to AVOID:

"blurry, low quality, distorted, watermark, text, artifacts"

Tips for 360° Panoramas

• Describe the ENVIRONMENT, not a single object
• Think in terms of "being inside" the scene
• Include directional variety: "mountains to the north, ocean to the south"
• Mention lighting that wraps around: "golden hour", "overcast sky"
• Avoid references to specific camera angles ("looking up", "close-up")

Prompt Themes

Reusable prompt themes let you save and apply commonly used prompt text: - Browse saved themes - Create and update custom themes - Apply a theme as a starting point, then edit the resulting prompt

Prompt Presets

Per-model recommended prompt/parameter combinations: - Curated by the 360 Hextile team for each processor - Include prompt text, negative prompt, guidance scale, steps - One-click application for quick starting points

AI Prompt Transform

Enhance your prompts using AI language models: - Type a simple idea and get an expanded, detailed prompt - Actions include generate, enhance, expand, contract, optimize, style, randomize, generate directions, variation, modify, and describe - Model-aware — the engine knows which pipeline you've selected and tailors phrasing to its text encoder and token budget (comma-separated tags for CLIP models like SDXL; natural-language sentences for T5 models like SD 3.5 and FLUX), so you get the right prompt language for the right model - Supports Claude and OpenAI integration — bring your own API key (encrypted local storage). This is the one optional feature that sends prompt text to an external provider; everything else runs locally.

Auto-Generate Directional Prompts

Let the engine write your cardinal/directional prompts instead of typing each one: - From your global prompt — describe the overall scene once, and the AI writes a distinct prompt for each direction (e.g. what a viewer sees looking north, east, south, west, up, down) - One click fills every direction in the active interpolation mode (6-Cardinal, 4-Cardinal, 3-Horizon, or 2-Point Gradient) - Directional prompts are appended per-tile to your global prompt and blended seamlessly across the sphere

Describe an Image (Local Vision Analysis)

Already have a panorama? Let the app look at it and write the prompts for you — runs entirely on your GPU, no API key required: - Choose a vision model: Florence-2 (fast/quality, ~1.2 GB VRAM) or Qwen-VL (best quality, ~11 GB VRAM) - The image is split into its six cubemap faces; each face is analyzed and turned into a per-direction prompt - Returns a populated global prompt, negative prompt, and all six cardinal prompts, plus a suggested theme name — ready to edit or render

Tiles (Geometry-Driven Directional Prompts)

Sample an arbitrary set of directions from any polyhedron, let Qwen2.5-VL describe each one, and get N aimed {yaw, pitch, prompt} points that blend seamlessly across the sphere — no typing required:

• Pick a sampling solid — choose any polyhedron from the 14 available shapes (Cube=6 faces, Dodecahedron=12, Icosahedron=20, up to Goldberg (3,0)=92). Default is Dodecahedron for a fast but expressive 12-direction sweep.
• Generate — the app perspective-crops your equirectangular input at each face-center direction, runs the local Qwen2.5-VL vision model on every crop in one batched session, and writes N directional prompts. No API key required — everything runs on your GPU.
• Editable point list — every generated {yaw, pitch, prompt} point is editable and deletable, same as Custom mode. Add or remove directions freely.
• Never stale on offset — the image offset (pan/rotate) is applied live at render time, so changing your view angle never invalidates the direction set. Regenerate only when you load a different input image.
• Persistent describe cache — content-addressed by image hash + settings. Describe the same image with the same polyhedron again and the cached points return instantly — no Qwen reload.
• Feeds Semantic Blend — when enabled, the N directional prompts blend in CLIP embedding space via spherical_mean for a spatially coherent "baked latent field" across the entire sphere. On families without semantic blend, degrades automatically to weighted text concatenation.
• Works with all diffusion pipelines (SDXL, SD3.5, SD2.1, SD1.5, FLUX, Z-Image Turbo, Qwen-Image).

Prompt Interpolation

For sequence workflows, interpolate between prompts over time: - 5 interpolation modes (linear, easing, cubic, etc.) - Smooth transitions between different scenes - Multi-prompt keyframe support

Token Counting

Real-time token estimate displayed as you type: - Shows how many tokens your prompt uses - Helps stay within model limits - Different processors have different token limits