ComfyUI Extension: ComfyUI · Egregora: Divide & Enhance

Authored by lucasgattas

Created

Updated

1 stars

Egregora: Divide & Enhance is a small suite of custom nodes that help you split, enhance, and recombine images, plus a clean SDXL prompt mixer that keeps things simple while staying robust with lot´s of customization.

Custom Nodes (0)

    README

    <center>

    ComfyUI · Egregora: Divide & Enhance 🧩🚀

    ⚡A set of ComfyUI nodes inspired by the Divide & Conquer algorithm, designed to split, enhance, and recombine images for high‑quality upscaling, with prompt mixing and analysis tools for cleaner, sharper results⚡

    <img width="1390" height="619" alt="image" src="https://github.com/user-attachments/assets/db49f88f-c6d4-4b74-899a-eb7e7edab5bb" /> </center>

    ✨ What is this?

    Egregora: Divide & Enhance is a small suite of custom nodes that help you split, enhance, and recombine images, plus a clean SDXL prompt mixer that keeps things simple while staying robust with lots of customization.

    Inspired by Steudio’s Divide & Conquer nodes, adapted and refactored to fit a streamlined upscaling workflow. 🧠✂️🧵

    Why it’s special

    • Deterministic plan: the same coordinate generator drives both Divide and Combine, so tiles align 1:1.
    • Seam‑aware recomposition: a controllable edge‑feather transparency (per‑tile) hides straight borders without creating gaps.
    • Order you can trust: linear, serpentine, spiral (in/out), and content/dependency‑aware orders for stable, reproducible passes.
    • Preview before compute: a visual overlay shows grid, overlaps, order, and feather zones.
    • Minimal prompts, strong conditioning: a turbo prompt node that just works with SDXL (with blacklist support).

    🧱 Included Nodes

    🚀 Egregora Turbo Prompt

    A minimal, predictable prompt builder:

    • Positive = caption_text + global_positive_prompt (plain concatenation; no weights).
    • Negative = global_negative_prompt only.
    • Blacklist (optional): removes whole‑word, case‑insensitive terms from the positive text before encoding.
    • Outputs SDXL‑compatible CONDITIONING with proper pooled output.

    Typical wiring

    • caption_text ← captioner (e.g., Florence2)
    • global_positive_prompt ← styles/quality you always want
    • global_negative_prompt ← artifacts to avoid
    • blacklist_words ← optional removals

    🧠 Egregora Algorithm

    Planner for Divide‑and‑Enhance upscaling.

    What it does

    • Computes an upscaled canvas that meets your minimum scale factor.
    • Derives a grid of tiles from your tile width/height and overlap.
    • Supports multiple tile orders: linear, serpentine, spiral_outward, spiral_inward, content_aware, and dependency_optimized.
    • Uses the same coordinate logic consumed by Divide & Select and Combine, ensuring perfect positional consistency.

    Why it matters

    • Prevents off‑by‑one and drift errors between split and merge phases.
    • Gives you predictable coverage for denoising/upscaling passes.

    ✂️ Egregora Divide & Select

    Splits an image according to the planner.

    Highlights

    • Emits all tiles or lets you select a specific tile for targeted enhancement.
    • Uses the exact coordinates and order from Egregora Algorithm.

    🔗 Egregora Combine

    Merges processed tiles back into a seamless image using order‑dependent alpha‑over with a per‑tile edge‑feather mask.

    How the seam fix works

    • Each tile gets an analytic feather that fades from 0 → 1 over a user‑controlled feather_size.
    • Feather is applied only on internal sides (outer canvas edges stay solid).
    • Effective feather is clamped to the tile overlap, so you don’t create holes.
    • Because the same coordinates are used as in Divide/Algorithm, the blend lands exactly where it should.

    Tips

    • Start with feather_size between 8–64 px; keep it ≤ min(overlap_x, overlap_y).
    • Use Preview to visualize feather bands vs. overlaps.

    👁️ Egregora Preview

    Visual overlay to preview the plan before you spend compute:

    • Draws tile boundaries, overlap zones, processing order, and feather bands for sanity checks.

    🔍 Egregora Content Analysis (optional)

    Analyzes gradient complexity to suggest:

    • Feather size (more for complex scenes),
    • Overlap adjustments, and
    • A recommended blending method mindset (feather‑first, then overlap).

    📦 Installation

    1. Clone into ComfyUI’s custom nodes folder:
    cd ComfyUI/custom_nodes
git clone https://github.com/lucasgattas/comfyui-egregora-divide-and-enhance.git
    1. Ensure the folder contains __init__.py and egregora_divide_and_enhance.py.
    2. Restart ComfyUI. Nodes appear under Egregora/.

    Note: You can rename the Python file; the package __init__.py controls what gets imported.

    🛠️ Minimal Workflow (SDXL)

    Goal: Use Turbo Prompt for text conditioning and keep the rest of your graph simple.

    1. Load Model (SDXL) → CLIP

    2. Egregora Turbo Prompt

      • caption_text = your main description
      • global_positive_prompt = secondary style or theme
      • blacklist_words = clean words from positive prompt (caption and global)

    3. KSampler (Euler/Karras or your favorite)

      • Positive/Negative from Turbo Prompt, model & latent as usual.

    4. VAE Decode → Preview / Save

    Upscaling (Divide & Enhance): Use Algorithm → Divide & Select → [process tiles] → Combine. Start with moderate tile sizes and overlap; preview first.

    📥 Example Workflow (.json)

    <img width="1130" height="742" alt="Captura de tela 2025-08-27 220324" src="https://github.com/user-attachments/assets/91735827-b882-45d7-bf17-d90fd23ed100" />
    • Download: examples/divide_and_enhance_example_workflow.json (included in this repo).
    • Import: ComfyUI → Queue (☰) → Load → pick the JSON, or just drag it inside.

    Prefer a one‑click experience? Run the improved, full tuned upscaler on the cloud with the best settings at (https://egregoralabs.com).

    🎚️ Practical Tips

    • Keep prompts concise; long texts dilute signal.
    • For realism vs. cartoon conflicts, add targeted negatives (e.g., cartoon, plush, chibi).
    • For seams: increase overlap a step or raise feather_size (stay ≤ overlap). Use Preview to inspect.
    • Large tiles reduce passes but increase VRAM — balance tile size vs. overlap for your GPU.

    🧪 Troubleshooting

    • Flat/grey outputs → raise steps slightly or increase the stronger slider; ensure the CLIP/model are SDXL compatible.
    • Repetitions/Grid looks → verify no external tiling patches are active; reduce extreme overlaps or tile size if using Divide/Combine.
    • Shape/key errors → update to the latest version; Turbo Prompt sets pooled output; make sure you pass a valid LATENT.

    🧾 Folder Structure

    comfyui-egregora-divide-and-enhance/
├─ __init__.py
├─ egregora_divide_and_enhance.py
├─ README.md  ← you are here
├─ examples/
└─  └─ divide_and_enhance_example_workflow.json

    🙌 Credits

    • Inspired by Steudio’s Divide & Conquer node suite and community best practices around SDXL conditioning and tiling.
    • Built for clarity: minimal controls, sensible defaults, strong integration with ComfyUI.

    📜 License

    GPL‑3.0 — see LICENSE.