comfyui-egregora-tiled

✨Tiled Regional Prompting + Model-level Tiled Diffusion for ComfyUI

---

🚀 Features

🔹 Image Tile Split

• Flexible grid with overlap (row-major or spiral ordering).
• Emits a shared grid_json used by the other nodes so everything stays in sync.

🔹 Tiled Regional Prompt

• Per-tile positive/negative prompts from lists or text.
• Global text (prepend/append/replace), blacklist filtering, and CLIP-encode caching.
• Validates and snaps each region to the latent grid to avoid off-by-one artifacts.

🔹 Egregora Magick Diffusion (Model)

• Wraps your model with tiled diffusion (Mixture/Multi/Spot Diffusion).
• Accepts the same grid_json (wildcard input) and auto-overrides tile sizes/overlap so your prompt regions and tiling match.
• Falls back gracefully if the external tiled module isn’t present.

✅ Works with SDXL (f=4) and SD1/2 (f=8) automatically (via scale/latent alignment). ✅ No extra pip installs.

---

📦 Installation

1) Via ComfyUI Manager (recommended)

1. Open ComfyUI → Manager.
2. Select Install from URL and paste this repo URL.
3. Click Install and then Reload ComfyUI.

2) Manual Installation

cd ComfyUI/custom_nodes
git clone https://github.com/lucasgattas/comfyui-egregora-tiled.git
# restart ComfyUI

---

🧩 Nodes

🖼️ Image Tile Split (Egregora)

Splits an IMAGE (B,H,W,C) into overlapping tiles and returns:

• tiles_batch: concatenated tiles
• grid_json: grid metadata for downstream nodes

---

✍️ Tiled Regional Prompt (Egregora)

Builds per-tile CONDITIONING (positive & negative) from lists or text. Supports:

• global_positive, global_negative
• blacklist_words
• combine_mode: prepend | append | replace
• CLIP text-encode cache for speed
• Optional bounds validation/snap to latent grid

Inputs: clip, grid_json (wildcard), prompt_list (wildcard) Outputs: positive, negative

---

🧪 Egregora Magick Diffusion (Model)

Model patcher that enables tiled diffusion and keeps tile sizing in lockstep with your grid.

• Methods: Mixture of Diffusers, MultiDiffusion, SpotDiffusion
• Reads grid_json (wildcard) and overrides tile_width, tile_height, tile_overlap automatically
• tile_batch_size to batch UNet calls when VRAM allows

Input: model (+ optional grid_json) Output: model (tiled-enabled)

---

⚡ Quick Start

1. Resize your input to the target resolution.
2. Image Tile Split → set tile_w, tile_h, overlap (e.g., 512/512/96–128).
3. Tiled Regional Prompt → connect the same grid_json and your clip; fill prompt_list (multi-line or comma-separated).
4. Egregora Magick Diffusion (Model) → connect the same grid_json so tile sizes/overlap match your prompt regions.
5. Sampler → use the patched model and the conditioning from the regional prompt.
6. VAE Decode → decode as usual (no special tiled decoder needed with this setup).

---

💡 Tips

• Larger tiles + smaller overlap = faster; more overlap = smoother seams (tune for your model).
• Reuse identical phrases across tiles to benefit from CLIP encode caching.
• Increase tile_batch_size if VRAM allows to amortize UNet overhead.

---

✅ Compatibility

• ComfyUI latest stable.
• Tested with SDXL and SD1.5.
• No additional Python packages required.

---

🗺️ What Changed (compared to earlier versions)

• Removed legacy Latent Tile Split and VAE Decode From Tiles sections from the workflow.
• Added the renamed Egregora Magick Diffusion (Model) node and streamlined the 3-node pipeline: Image Tile Split → Tiled Regional Prompt → Egregora Magick Diffusion (Model).