DAAM Extension for Stable Diffusion Web UI / Forge

This extension is a WebUI script port of DAAM with compatibility fixes for recent WebUI and Forge backends.

Features

• Attention heatmaps for comma-separated target phrases.
• BREAK is also treated as a separator in attention text input.
• Dynamic Prompts-aware term resolution for attention targets:
  • Nested variants, e.g. {{bunnygirl outfit|cat girl}|school girl}
  • Weighted/ranged variants, e.g. {0.2::red eyes|0.8::blue eyes}, {1-2$$ and $$a|b|c}
  • Sampler-prefixed variants, e.g. {~hands up|arms crossed}, {@a|b}
  • Wildcards and wildcard globbing, e.g. __artists__, __artists*__
  • Uses Dynamic Prompts parser + wildcard manager when available, with legacy fallback resolver
• Explicit Enable DAAM toggle (on/off like other always-on scripts).
• Time-focus controls with explicit on/off:
  • Enable time focus
  • modes: Disabled (default), All, Early, Mid, Late, Triplet
• Triplet mode renders Early + Mid + Late in one run.
• Optional diagnostics JSON (*_daam_diag.json) with per-term match status and reason codes.
• Prompt influence modes:
  • Positive: visualize conditional prompt attention.
  • Negative: visualize unconditional/negative prompt attention.
  • Delta (Pos-Neg): max(0, positive - negative) (suppression-aware positive remainder).
  • Signed Delta (Pos-Neg): signed positive - negative normalized to [-1, 1] and mapped to [0, 1].
  • Abs Delta (|Pos-Neg|): absolute difference between positive and negative attention.
• SDXL-compatible prompt token handling.
• Forge-compatible UNet and text-encoder resolution.
• Forge Anima support, including its JointTextEncoder/Qwen + T5 context and per-block heatmaps.
• Grid output mode for multiple attention terms.

Install

Clone into your WebUI/Forge extension directory:

Usage

1. Open txt2img or img2img.
2. Expand Attention Heatmap.
3. Enable Enable DAAM.
4. Enter target terms in Attention texts for visualization (comma separated).
5. Generate.

The extension creates blended heatmap images and, if enabled, a grid_daam-*.png in the grid output directory.

Notes:

• Multi-word phrases are matched as one sequence (for example white dress).
• You can separate attention targets with , or BREAK.
• For Dynamic Prompts, DAAM resolves attention templates against the resolved per-image prompt text.
• You can put Dynamic Prompts syntax directly in the DAAM attention box (variants/wildcards); DAAM will try to resolve it to the concrete token used in that image.
• Variable/Jinja-heavy templates are handled best-effort. If a term cannot be uniquely resolved, diagnostics will show a miss reason.
• DAAM normalizes extra-network tags (for example LoRA) for internal token mapping, so LoRA position (start/middle/end) should not change heatmap term matching.
• DAAM heatmaps are generated in the image-save callback (on_before_image_saved).
• In WebUI, keep Always save all generated images enabled.
• For API generation, set "save_images": true; otherwise no DAAM heatmaps are produced.
• All means one aggregated heatmap over all denoising steps.
• Triplet means three phase heatmaps (Early, Mid, Late) in one generation.
• For Triplet, using Use grid (output to grid dir) is recommended for easier comparison.
• Delta (Pos-Neg) can look empty if a term is only present in negative attention (pos <= neg everywhere after clamping).
• Use Signed Delta or Abs Delta when you want to inspect suppression/conflict directly instead of only positive remainder.

Forge Notes

Hires Fix / Upscaling

DAAM works with Hires Fix and non-latent upscalers (for example 4x-UltraSharp).

For API calls on some Forge builds, include:

"hr_additional_modules": []

Without this field, Forge can raise:

TypeError: argument of type 'NoneType' is not iterable

This error originates in Forge processing, not in DAAM logic itself.

ADetailer

ADetailer performs an internal pre-pass postprocess callback before final image saving.
DAAM now defers final cleanup for that dummy callback so heatmaps remain available for the final output save stage.

Current behavior:

• DAAM heatmaps are generated for the main generation pass.
• ADetailer can still change the final image composition after the main pass, so heatmaps should be interpreted as prompt attention guidance, not strict post-inpaint attribution.

Batch Size / Batch Count

DAAM now supports Forge runs with both:

• batch_size > 1
• n_iter (batch count) > 1

The extension no longer depends on Batch pos PNG metadata (which may be absent on Forge).
Instead, it resolves per-image batch positions from processing state and seed/filename fallback logic, and handles compact Forge cond_or_uncond layouts in attention tracing. It also keeps per-batch prompt analyzers so variable prompts in one batch resolve terms correctly.

Output Folder Layout

To avoid missing heatmaps, keep sample and grid output folders separate.

Recommended:

• outdir_txt2img_samples = outputs/txt2img-images
• outdir_img2img_samples = outputs/img2img-images
• outdir_txt2img_grids = outputs/txt2img-grids
• outdir_img2img_grids = outputs/img2img-grids

Avoid setting grid dirs to a parent folder that also contains sample dirs (for example outdir_txt2img_grids = outputs with samples in outputs/*), because this can cause grid/sample misclassification in legacy configs.

Sample

• Prompt: A photo of a cute cat wearing sunglasses relaxing on a beach
• Attention text: cat, sunglasses, beach
• Output: original + one heatmap per attention term