ComfyUI-KrakenTools

🦑 ComfyUI-KrakenTools

High-utility nodes for ComfyUI with a focus on Flux 1 Dev workflows and Ultimate SD Upscale enhancement loops.

Why this exists: when your “final” resolution is relatively small (e.g., 736×1344 portraits), running no real upscale in Ultimate SD Upscale (USDU) often doesn’t add detail. The Kraken approach enforces a true enlarge pass (e.g., 1.5×–2×), then downscales back to your exact target with high-quality filtering — preserving micro-detail and cleaning seams.

✨ Nodes

• 🦑 Kraken Flux Empty Latent Image (KrakenFluxEmptyLatentImage)

  • Builds a Flux-style latent and exposes canonical preset + resolution strings for downstream nodes.
  • Outputs a small UI preview so you always see the intended canvas size.

• 🦑 Kraken Upscale & Tile Calc (KrakenUpscaleTileCalc)

  • Computes a uniform upscale factor to cover your target, aligns to model-friendly multiples, and suggests tile sizes + overlaps for USDU.
  • New: min_upscale_policy (auto/off/1.5x/2.0x/custom) guarantees USDU actually upscales, even when target==source.

• 🦑 Kraken Resolution Helper (Exact) (KrakenResolutionHelper)

  • Exact-size finisher: keep proportion (contain + pad), pad (with anchors), fill/crop (cover), or stretch.
  • Downscales without cropping and guarantees final WxH. With allow_upscale=False, it will never enlarge in the final step.

🛠 Installation

1. Locate your ComfyUI custom_nodes directory.
2. Clone or copy this repo into it:

   cd /path/to/ComfyUI/custom_nodes
   git clone https://github.com/<your-user>/ComfyUI-KrakenTools.git
   

3. Restart ComfyUI.

Requirements

• ComfyUI (latest)
• Python deps: torch, numpy, Pillow (bundled in ComfyUI)
• Ultimate SD Upscale node pack (for USDU) installed separately

🚀 Usage (portrait example 736×1344)

1. Flux latent → add KrakenFluxEmptyLatentImage and pick your preset (e.g., 736×1344).
   Outputs: LATENT, resolution (e.g., "736 x 1344"), preset_out (e.g., "736x1344"), width/height, preview

2. Upscale planning → add KrakenUpscaleTileCalc

   • resolution_in ← latent.resolution
   • preset_in ← latent.preset_out
   • min_upscale_policy = auto (2× for small canvases), or force "2.0x"

3. USDU → add Ultimate SD Upscale (regular node)

   • upscale_by ← tile calc.upscale_by
   • Tile/overlap settings ← tile calc outputs
   • Denoise ~0.2–0.35

4. Final sizing → add KrakenResolutionHelper

   • image ← USDU output
   • preset_in ← tile calc.target_resolution (or latent.preset_out)
   • mode = keep proportion, interpolation = lanczos, allow_upscale = False
   • Output is exact 736 × 1344 with extra micro-detail

🧭 Contributing

See CONTRIBUTING.md.

🪪 License

MIT © 2025 The Kraken — X: @KrakenUnbound · YouTube: /@KrakenUnbound