Prompts
Write, reuse, describe and refine the prompts that shape a panorama
Last updated: 2026-10-27
Prompts guide the selected model. Start with the environment, materials, lighting and mood; compare output rather than adding generic quality words indefinitely. For example: “a quiet mountain lake, pine forest, distant snow peaks, soft evening light.”
Writing Effective Prompts
Use the model's recommended wording and token budget. Negative prompts apply only where supported; a field from a saved Workflow may not be used by a different pipeline. Style/quality presets and LoRA words can add text beyond the global prompt. See Parameters.
Prompt Themes
Reusable environments are Worlds. Open the toolbar's Worlds browser to load, append or prepend text, favorite entries and save/export your own. Worlds explains what is replaced and how Load Optimized differs from ordinary Load.
Prompt Presets
Quality/negative preset fragments are model-specific helpers, not a full Workflow or a substitute for inspecting the rendered result. Style pills add aesthetic text. LoRA-provided style wording and trigger words have their own controls; avoid duplicating the same wording by hand without checking the final setup.
AI Prompt Transform
Generate creates prompts from an idea; Modify applies your instruction to existing targeted text. Refinement actions include Enhance, Optimize, Expand and Contract. Randomization/style actions appear where that toolbar supports them. Spot/tile surfaces show a smaller relevant set rather than the full main toolbar.
Check Acting on and the selected/checked tiles before applying. A scoped action should affect its current targets, not every prompt. Read the changed text, then use the offered Undo if it missed your intent.
AI text actions need a configured Claude, OpenAI or xAI provider/key and send the needed text/context to that provider. Choose an actually discovered model; provider charges and limits can apply. This is one online feature among several, not the app's only network use. See Privacy.
Auto-Generate Directional Prompts
Use Generate/Modify with the intended tile targets, or author aimed prompts yourself. Current directional work uses a list of prompt tiles with positions and influence, not a required choice of six-cardinal/four-cardinal interpolation modes. Review the aim and text before rendering.
Describe an Image (Local Vision Analysis)
Describe uses the local Qwen vision model rather than your text-provider key. It still needs installed vision weights and GPU readiness. Select the intended input/scope and sampling settings before starting; generated text is a proposal to inspect, not a guarantee that every visible object was understood.
Use the offered tile/polyhedron sampling for multiple aimed descriptions, and review whether the action also updates the global prompt. Selected targets and current surface determine what is replaced. Do not promise that every Describe action always returns exactly six directions and a negative prompt. A cancelled or failed description should be checked before trying again.
Tiles (Geometry-Driven Directional Prompts)
Prompt Tiles let you place local descriptions around the sphere. Reach controls how widely an aimed prompt influences nearby directions; Influence Cutoff suppresses small contributions. Auto uses the current model's settings. FOV controls a sampled description crop; it is not another name for Reach. Model-specific Semantic Blend, where available, changes how influences combine; it is not an output-quality guarantee.
A tile's own prompt can replace the global prompt; include-global options deliberately combine them where offered. Check scope and preview before a full pass. See Directional Prompts for aiming and influence details.
Prompt Interpolation
Spatial blending between aimed directions is different from changing prompts over a timed animation workflow. Advanced temporal workflows have their own availability and controls; do not infer them from a spatial prompt list. See Sequences.
Token Counting
The toolbar's counter describes Global Prompt and may be estimated. Prompt Tiles, styles and quality presets can add tokens per rendered tile. Staying under the global count alone does not prove the final composed text fits every model. Shorten or optimize for the selected model, then compare a small render.