StepX — Attention Explorer

Compute heatmap for specific pixels

Pixel Heatmap shows which regions of the image are attended to by specific pixels.

🎯 How to Use:

1. Click on Image: Click on any pixel in the generated image above, and the system will automatically identify and generate the heatmap (Recommended)

1. Manual Input: You can also manually enter pixel IDs (row-major order, 0-4095 for SD v2-base):

• Single pixel: 100

• Range: 0-1023

• List: 0,1,2,3

Compute heatmap for bounding box

BBox Heatmap shows the average attention heatmap for all pixels within a bounding box.

🎯 How to Use:

1. Click on Image: Click twice on the generated image above to define the bounding box (first click: top-left corner, second click: bottom-right corner)

1. Manual Input: You can also manually enter coordinates in latent space (0-63 for SD v2-base):

Compute heatmap for polygon contour

Contour Heatmap uses polygon contours to define regions (supports irregular shapes).

🎯 How to Use:

1. Click on Image: Click multiple times on the generated image above to add contour points (at least 3 points required)

1. Manual Input: You can also manually enter contour points in image coordinates:

• Format 1: x1,y1 x2,y2 x3,y3 ... (e.g., 0,0 256,0 256,256 0,256)

• Format 2: x1 y1 x2 y2 x3 y3 ... (e.g., 0 0 256 0 256 256 0 256)

Compute diffused heatmap from a pixel

Pixel Diffused Heatmap starts from a single pixel and iteratively diffuses attention, gradually enhancing related regions.

Suitable for discovering semantically related regions of pixels.

🎯 How to Use:

1. Click on Image: Click on any pixel in the generated image above to select the starting pixel (Recommended)

1. Manual Input: You can also manually enter the starting pixel ID

Text-Guided Object Detection (TITAN Workflow)

Combines DAAM (Cross-Attention) and DAAM-I2I (Self-Attention) for text-guided object detection and segmentation.

Important: The detection result will be displayed on the DAAM-I2I generated image, not the DAAM image.

How It Works:

1. DAAM tab: Generate an image with a prompt. This computes cross-attention (text-to-image attention) and creates word heatmaps.

1. DAAM-I2I tab: Generate an image (can use the same or different prompt). This computes self-attention (image-to-image attention).

1. Detection: Enter a word to detect. The system will:

• Extract the word's heatmap from DAAM (based on DAAM prompt and image)

• Use this heatmap as guidance for DAAM-I2I self-attention (based on DAAM-I2I prompt and image)

• Display the detection result on the DAAM-I2I image

⚠️ Important Notes:

• Same prompts: If both tabs use the same prompt, detection works best as the word heatmap matches the image content.

• Different prompts: If prompts differ, the word you're detecting must exist in the DAAM prompt. The detection will still be shown on the DAAM-I2I image, but results may be unexpected if:

• The word doesn't exist in the DAAM-I2I prompt

• The word refers to different objects in the two prompts

• The images are completely different