Denrakeiw Nodes

Comprehensive ComfyUI custom nodes for image generation, latent space manipulation, Flux LayerDiffuse transparency, and multi-image composition.

🎨 Core Nodes

Color Generator

Generates solid color images with 130+ predefined colors.

• Input: Width/height (1-8192px), color selection
• Output: Image, color name, hex code

📁 Load Image Sequence

Sequential image loading from folders with state management.

• Features: Auto-increment, sorting options, loop mode, error handling
• Output: Current image, filename, index, total count
• Advantage: Low memory usage vs batch loading

🖼️ Multi-Image Aspect Ratio Composer

Advanced multi-image composition with intelligent layouts.

🎛️ Dynamic Controls:

• Input Count: 1-8 images with dynamic UI updates
• Aspect Ratios: 1:1, 4:3, 16:9, 9:16, 21:9, 3:2, 5:4, 2:1, etc.
• Megapixels: 0.5-32 MP with auto-calculation
• Layouts: Horizontal, Vertical, Smart Grid, Classic Grid

🎭 Face Detection:

• Haar Cascade: Fast CPU-based detection
• DNN Face: Advanced deep learning detection
• Smart Cropping: Centers images around detected faces
• Confidence: Adjustable detection threshold (1.1-3.0)

🎯 Smart Grid Algorithm: Optimizes layout based on target aspect ratio:

• 8 images + 16:9: [4,4] layout (2 rows)
• 8 images + 1:1: [3,3,2] layout (3 rows)
• 8 images + 9:16: [2,2,2,2] layout (4 rows)

---

🎨 Latent Space Tools

ℹ️ The latent nodes below are also available as standalone packs (Latent_Nodes and WAN_NN_Latent_Upscale). To allow both packs to coexist in one ComfyUI install without node-ID conflicts, the integrated versions here use an internal _DRE suffix and a *DRE tag in the display name (e.g. 🎨 Latent Color Match *DRE).

🎨 Latent Color Match *DRE

Advanced color matching directly in latent space - 10x faster than image-based methods.

• Methods: LAB, YCbCr, LUV, YUV, XYZ, RGB (kornia) + advanced algorithms (color-matcher)
• Benefits: No VAE encoding/decoding, 50% less VRAM, no quality loss
• Controls: Factor strength (0.0-3.0), batch processing, GPU acceleration

🎛️ Latent Image Adjust *DRE

Complete adjustment suite in latent space with kornia integration.

• Controls: Brightness (-1.0 to 1.0), Contrast (0.0 to 3.0), Hue (±180°), Saturation (0.0-3.0), Sharpness (0.0-3.0)
• Features: Direct latent manipulation, memory efficient, anti-aliasing

🚀 Universal NN Latent Upscale *DRE

Neural network-based latent upscaling (from the WAN_NN_Latent_Upscale repo, based on Ttl's ComfyUi_NNLatentUpscale).

• Models: SD1.5, SDXL, Flux, and Wan2.2 (bundled .pt files in models/)
• Scale: Configurable 1.0x–2.0x, automatic model detection/loading
• Performance: Direct latent processing without VAE decode/encode cycles

Universal Latent Upscaler

Neural network-based latent upscaling (custom DenRakEiw V2.0 auto-detecting 4ch/16ch).

• Models: Supports multiple architectures with automatic detection
• Performance: Direct latent processing without VAE decode/encode cycles

---

🔥 Flux LayerDiffuse (Transparent Images)

Complete transparent image generation system for Flux models.

Core Nodes:

• 🔧 Standalone Loader: Loads TransparentVAE.pth model
• 🔍 Decoder (Simple): Converts latents to transparent RGBA images
• 💾 Save Transparent: Saves PNG with proper alpha channel
• 👁️ Preview Transparent: Multi-background preview (checkerboard/white/black)
• 📊 Transparent Info: Alpha channel analysis and statistics

Utility Nodes:

• 🔍 Conditioning Inspector: Verifies Flux tensor compatibility
• 🔧 Conditioning Fix: Converts SD to Flux format (if needed)
• ⭕ Empty Conditioning: Proper negative conditioning for Flux
• ℹ️ Info: Model requirements and download links

Quick Workflow:

1. Models: UNETLoader (Flux) + LoraLoader (layerlora.safetensors, strength 1.0)
2. Text: DualCLIPLoader (type="flux") + CLIPTextEncode
3. Generate: EmptySD3LatentImage + KSampler
4. Transparent: Standalone Loader + Decoder + Save/Preview

Example Prompts: "transparent glass bottle", "crystal wine glass", "clear ice cube"

---

🎬 Flux 3 API (BFL)

Two nodes for the BFL Flux 3 video API and an LLM-powered prompt generator.

🎬 Flux 3 Video (API)

Generates video via POST https://api.bfl.ai/v1/flux-3-video with synchronized audio. Four modes: t2v (text→video), i2v (image→video with keyframes), v2v (video continuation), draft_enhance (final-render a prior draft). Up to full-HD, 5–20 seconds. Async with retries, backoff, and parallel execution. (Docs)

🎬 Flux 3 Openrouter Prompt

Turns a plain-words idea into a complete, structured FLUX 3 prompt using an OpenRouter LLM. The model dropdown is filled live from the OpenRouter API. Optional reference images for vision-capable models. Output is just the prompt — no variants, no metadata — ready to paste into the Video node's prompt input.

Skills: Two prompting skills are pre-installed in the skills/ folder:

• Flux3Director — full structured prompting skill (all modes, timing, camera vocabulary, tag system)
• Flux3Director4Discord — compressed for the FLUX3 Discord bot's 2,000-character limit

Drop your own .md files into skills/ and refresh ComfyUI — they appear in the dropdown automatically.

Setup: Add to .env:

BFL_API_KEY=bfl_...
OPENROUTER_API_KEY=sk-or-...

---

📁 Installation & Requirements

Installation:

1. Clone to ComfyUI/custom_nodes/denrakeiw_nodes/
2. Install dependencies: pip install -r requirements.txt
3. Restart ComfyUI

Required Models (Flux LayerDiffuse):

• Flux: flux1-dev-fp8.safetensors → models/diffusion_models/
• LoRA: layerlora.safetensors → models/loras/ (included)
• TransparentVAE: TransparentVAE.pth → models/vae/ (included)
• CLIP: clip_l.safetensors, t5xxl_fp8_e4m3fn.safetensors → models/clip/
• VAE: ae.safetensors → models/vae/

Download from: Hugging Face - Flux, LayerDiffuse

Dependencies:

pip install kornia>=0.6.0 color-matcher>=0.2.0 diffusers==0.32.2 safetensors transformers peft opencv-python

🚀 Quick Usage

Multi-Image Composer:

1. Set Input Count (1-8) → Click Update Inputs
2. Choose Aspect Ratio and Megapixels
3. Select Layout: Smart Grid (recommended) or Horizontal/Vertical
4. Enable Face Detection for portraits (Haar Cascade/DNN)
5. Connect images → Execute

Latent Color Matching:

Image → VAE Encode → Latent Color Match → VAE Decode → Save

Transparent Images (Flux):

UNETLoader + LoraLoader (layerlora.safetensors) → KSampler →
Flux LayerDiffuse Decoder → Save Transparent Image

🎯 Troubleshooting

Flux LayerDiffuse:

• Tensor errors: Use DualCLIPLoader type="flux" + Conditioning Inspector
• Glass generation: Normal behavior - reduce LoRA strength to 0.5-0.8
• No transparency: Use Save Transparent Image (not standard Save Image)

Multi-Image Composer:

• Missing images: Check all inputs connected + click "Update Inputs"
• Face detection slow: Disable for non-portrait images
• Memory issues: Reduce megapixel setting

🙏 Acknowledgments

Thanks to RedAIGC (Flux LayerDiffuse), lllyasviel (LayerDiffuse foundation), and lum3on (support & inspiration).