Directional Tiles

Alpha — available for testing and feedback. Controls and behavior may change.

What It Is

Directional Prompt Mode lets you assign different prompts to different viewing directions instead of one prompt for the whole image. It uses Qwen2.5-VL, a local vision model, to drive geometry-based prompting across polyhedra face-center directions.

Each polyhedron face points to a specific direction on the sphere. You assign a prompt per face, and the AI generates content matching each direction -- different sky, ground, and everything between.

How It Works

The equirectangular image is divided into directional zones based on polyhedra geometry:

1. Pick a polyhedron -- 14 shapes available, from Cube (6 faces) to Goldberg GP(3,0) (92 faces). More faces = finer directional control, slower generation.
2. Assign prompts per face -- each face maps to a {yaw, pitch, prompt} point. Editable, deletable, freely reorderable.
3. Prompts mix across the sphere -- between face centers, normalized Gaussian influence combines nearby prompt anchors.

Available shapes and their face counts:

PolyhedraTileControls

The Tiles list shows every face-direction tile. Each card displays its direction (yaw/pitch), assigned prompt, and Describe status. The same list is used for Prompt tiles and for Spot rows.

• Select -- click a tile's metadata bar or Target icon to load it into the editor. A single click does not move the 360 camera.
• Look at -- double-click the metadata bar or Target icon to point the 360 viewer at that tile. The view glides using your Camera smoothing setting. Zoom (FOV) does not change. Face-only tiles with no yaw/pitch do nothing on double-click. In the compact list, Enter on a selected posed row does the same.
• Aim overlay -- on the 360 HUD, the selected marker shows its current mix share. Show all adds every other marker whose share survives Cutoff. Right-click blank space to Add Spot Here; right-click a marker to Duplicate Spot or Delete Spot.
• PolyhedraPreview -- a rotatable 3D solid rendering showing all face center points. Hover or click faces to see their prompts.
• Editable point list -- every generated point is editable and deletable. Add or remove directions freely.

Qwen2.5-VL Describe

Click Describe on a tile to have Qwen2.5-VL analyze what is currently visible in that direction and suggest a prompt. No API key required -- everything runs locally on your GPU.

• How it works -- the app perspective-crops your equirectangular input at the face-center direction, runs Qwen2.5-VL on the crop, and writes a prompt describing the content.
• Batch describe -- use Describe All to run Qwen2.5-VL on every face in one batched session. All directional prompts populate at once.
• Describe cache -- results are content-addressed by SHA-256 hash of the input image and cached in an LRU cache of 256 entries. Describe the same image with the same polyhedron again and results return instantly -- no Qwen reload.
• Qwen2.5-VL is the only describe provider. Requires ~11 GB VRAM. Install it from the Pipelines page.

Gaussian prompt interpolation

Each render tile gets a normalized Gaussian mix based on angular distance and neighbouring anchor spacing. Reach sets the influence width; Auto Reach derives it from nearby interpolation anchors. Cutoff removes low-share prompts from the mix. This is not cosine weighting.

When Semantic Blend is enabled (Alpha, on supporting pipelines), directional prompts combine in CLIP embedding space via spherical_mean. Pipelines without semantic blend use weighted text concatenation.

Image Offset Interaction

Directional prompts stay anchored to absolute sphere directions. When you offset (pan/rotate) the image, the image moves under the fixed prompt grid:

• Your "sky" prompt always points up, regardless of where the image content sits.
• Your "ground" prompt always points down.
• Changing your view angle or panning the image never invalidates your directional prompt set.
• Regenerate prompts only when you load a different input image.

The image offset (yaw, pitch, roll) is applied live at render time. The prompt grid is world-space absolute; the image is camera-space relative.

Generate Flow

1. Pick a polyhedron -- select the shape that matches your desired directional granularity.
2. Place directional prompts -- type prompts per face, or leave some blank to inherit the global prompt.
3. Optional: Describe -- click Describe or Describe All to let Qwen2.5-VL analyze each direction and suggest prompts.
4. Click Generate -- the AI fills each direction according to its assigned prompt, blended smoothly across the sphere.
5. Review -- check the result in the viewer. Edit prompts and regenerate specific directions as needed.

Tile Ordering Strategies

15 strategies control the order in which tiles are generated. The order affects how adjacent tiles influence each other during processing:

Works with all visible diffusion pipelines (SDXL, SD3.5, SD2.1, SD1.5, FLUX, Z-Image Turbo, and Qwen-Image in Alpha builds).

Navigator FOV and prompt influence

When placing or inspecting directional points in the viewer/navigator:

• Look at a tile from the Tiles list (double-click, or compact Enter) to glide the 360 view to that direction without changing FOV.
• FOV changes the perspective crop used by Describe. It does not change prompt mixing.
• Reach controls Gaussian influence width. Cutoff removes small shares. Use those controls when a direction should affect more or less of the sphere.
• Place-and-describe workflows: drop a direction, write its prompt, then let auto-fill or AI transform expand the set — product language is Directional.

Upscale vocabulary (when chaining after directional work)

Upscaling has several workflow surfaces. On an open render, one Resize toolbar popover now contains both absolute Size destinations and native-relative Upscale factors; there is no separate Upscale toolbar button.

See Upscale.

Next Steps

• Prompt System -- Interpolation modes and prompt tools
• Interface Overview -- Where prompt tools live in the app shell
• Templates -- Choose the right template
• Pipelines -- Understand different AI models