Kraken Tools for ComfyUI

A collection of productivity custom nodes for ComfyUI, designed to streamline image generation workflows with smart defaults, advanced features, and quality-of-life improvements.

Installation

Via ComfyUI Manager (Recommended)

Search for "Kraken Tools" in ComfyUI Manager and click Install.

Manual Installation

cd ComfyUI/custom_nodes
git clone https://github.com/yourusername/kraken_tools.git
cd kraken_tools
pip install -r requirements.txt

Requirements

• ComfyUI
• Python 3.10+
• PyTorch 2.0+
• transformers >= 4.45.0 (for Kraken Unbound Prompt)
• Pillow
• requests (for CivitAI trigger word fetching)
• opencv-python (for image processing)
• scipy (for image processing)
• ollama (optional, for Kraken Ollama Prompt Chat)

---

Nodes Overview (15 Nodes)

Prompt & Text Nodes

🐙 Kraken Unbound Prompt

Advanced prompt builder with built-in vision model support for image-to-text captioning.

Features:

• Vision Mode: Use Qwen2-VL-2B-Instruct to generate prompts from images
• Prompt Enhancement: Automatically enhance prompts using AI
• Style Presets: Choose from photorealistic, cinematic, anime, manga, fantasy, sci-fi, and more
• Lighting Options: Soft, dramatic, cinematic, studio, natural, golden hour, neon, rim, backlighting
• Camera Settings: Lens type, f-stop, bokeh effect, DSLR style
• Persistent Prompt: Add text that always appears (e.g., "blue skies", "brand logo")
• Negative Prompt: Separate negative prompt output
• Position Control: Prepend or append style settings to your prompt
• VRAM Management: Control model loading/unloading with keep_alive_minutes

Inputs:

Input	Type	Description
prompt	STRING	Your base prompt text
use_image_as_source	BOOLEAN	Use vision model to caption an image
input_image	IMAGE	Image for vision mode
enhance_prompt	BOOLEAN	Use AI to enhance the prompt
enhancer_style	DROPDOWN	Modern, Classic Tags, Instructional, or WAN style
negative_prompt	STRING	Negative prompt text
persistent	STRING	Text always added to the prompt
style/lighting/camera_lens/f_stop	DROPDOWNS	Style settings
position	DROPDOWN	Prepend or append style settings
force_unload	BOOLEAN	Unload model after use to free VRAM

Outputs:

• positive_prompt (STRING): The final positive prompt
• negative_prompt (STRING): The negative prompt
• persistent (STRING): The persistent prompt (passthrough)

---

🐙 Kraken WAN Prompt Splitter

Split and style prompts for WAN video generation with comprehensive cinematic presets.

Features:

• Prompt Splitting: Split a single prompt into up to 5 segments using delimiters (---, ###, etc.)
• WAN Packs: Pre-configured style packs (Cinematic Natural, Moody Neon, Documentary Daylight, Epic Vista, Noir Classic)
• Shot Type: extreme close-up, close-up, medium shot, cowboy shot, wide shot, aerial, POV, etc.
• Lens Selection: 24mm to 200mm, anamorphic, tilt-shift, fisheye, macro
• Aperture: f/1.4 dreamy bokeh to f/16 deep focus
• Lighting: natural daylight, golden hour, neon mix, studio 3-point, film noir, candlelight, etc.
• Film Stocks: photoreal, portra-like, cinestill-like, black-and-white, vintage chrome, etc.
• Color Grading: warm amber teal, cool cyan steel, muted pastels, high contrast, etc.
• Composition: rule of thirds, centered symmetry, leading lines, dutch angle, etc.
• Environment: clear, fog, rain, snowfall, dust storm, haze
• Texture: ultra-detailed, fine film grain, clean minimal, gritty texture

Inputs:

Input	Type	Description
prompt	STRING	Main prompt (use --- to split)
persistent	STRING	Text added to all prompts
wan_pack	DROPDOWN	Pre-configured style pack
shot_type/lens/aperture/etc.	DROPDOWNS	Individual style controls
combine_persistent	BOOLEAN	Add persistent text to prompts
combine_style_presets	BOOLEAN	Add style settings to prompts

Outputs:

• Prompt 1 through Prompt 5 (STRING): Split and styled prompts

---

🐙 Kraken Ollama Prompt Chat

Connect to a local Ollama LLM instance for interactive prompt enhancement.

Features:

• Connect to any Ollama server
• Use any Ollama model (llama3.2, mistral, etc.)
• Interactive chat for refining prompts
• Specialized system prompt for AI art generation

Inputs:

Input	Type	Description
host	STRING	Ollama server URL (e.g., http://localhost:11434)
model	STRING	Model name (e.g., llama3.2)
action	DROPDOWN	Connect, Send Message, Clear Chat, Use Last Response
user_message	STRING	Your message to the LLM

Outputs:

• prompt_output (STRING): The generated prompt
• ai_response (STRING): Full AI response
• conversation_log (STRING): Chat history

---

Model Loading Nodes

🐙 Kraken Checkpoint Loader

Enhanced checkpoint loader with model tracking.

Features:

• Load checkpoints with optional dtype control (fp16, bf16, fp32)
• Output model name for metadata tracking
• Compute SHA-256 hash (first 10 chars) for version tracking

Outputs:

• MODEL, CLIP, VAE: Standard model outputs
• model_name (STRING): Filename of loaded checkpoint
• ckpt_sha10 (STRING): First 10 chars of SHA-256 hash

---

🐙 Kraken Dual CLIP Loader

Full-featured dual text encoder loader for modern models.

Features:

• Support for Flux, SDXL, SD3, and Hunyuan models
• Load CLIP-L + T5 text encoders
• Modes: dual, t5-only, clip-only
• Device placement control (auto, cuda, cpu)
• Precision override (fp16, bf16, fp32)
• Optional warmup for faster first inference

Outputs:

• clip (CLIP): Combined CLIP object
• diagnostics (STRING): Loading information

---

🐙 Kraken LoRA Loader (3)

Load up to 3 LoRAs with automatic trigger word fetching.

Features:

• Load 3 LoRAs simultaneously with individual enable toggles
• Per-LoRA strength control for both model and CLIP
• Per-LoRA CLIP skip setting
• CivitAI Integration: Automatically fetch trigger words from CivitAI
• API key stored securely in ~/Documents/ComfyUI/user/kraken_config.json
• Trigger word placement: prepend, append, or replace
• Caches trigger words locally for offline use

Inputs:

Input	Type	Description
model/clip	MODEL/CLIP	Input model and clip
lora_X_enabled	BOOLEAN	Enable/disable each LoRA
lora_X_file	DROPDOWN	LoRA file selection
lora_X_model_strength	FLOAT	Model weight (-2.0 to 2.0)
lora_X_clip_strength	FLOAT	CLIP weight (-2.0 to 2.0)
lora_X_clip_skip	INT	CLIP skip (0 = off)
lora_X_force_fetch	BOOLEAN	Re-fetch trigger words
fetch_triggers	BOOLEAN	Enable CivitAI fetching
placement	DROPDOWN	prepend, append, or replace

Outputs:

• model (MODEL), clip (CLIP): Modified model/clip
• prompt (STRING): Prompt with trigger words
• lora_names, triggers, tags (STRING_LIST): Metadata

---

Latent & Resolution Nodes

🐙 Kraken Empty Latent Image

Create empty latent images with preset aspect ratios.

Features:

• Megapixel selection: 0.5 to 3.0 MP or Custom
• Preset aspect ratios organized by category:
  • Vertical/Portrait (9:20, 9:16, 2:3, 3:4, 4:5)
  • Square (1:1)
  • Landscape (5:4, 4:3, 3:2, 16:10, 16:9)
  • Cinematic (1.85:1, 2:1, 21:9, 2.39:1)
  • Screen sizes (Android, iPhone, Full HD, 4K)
  • Social media sizes (Instagram, YouTube, Facebook)
• Custom aspect ratio input
• Divisible-by alignment (8, 16, 32, 64)

Outputs:

• latent (LATENT): Empty latent for generation
• width_out, height_out (INT): Dimensions
• resolution (STRING): "1024 x 1024" format
• preset_out (STRING): For connecting to upscale calculator

---

🐙 Kraken Resolution Helper

Ensures final output matches a specific target resolution. Batch-safe for video.

Features:

• Multiple modes: fill/crop, pad, keep proportion, stretch
• Interpolation: lanczos, bicubic, bilinear, nearest
• Anchor positioning for crop/pad (left/center/right, top/center/bottom)
• Upscale prevention option
• Works with batched images (video frames)

Inputs:

Input	Type	Description
image	IMAGE	Input image batch
mode	DROPDOWN	Resize mode
target_resolution_in	STRING	Target as "WxH"
preset_in	STRING	From Kraken Empty Latent

---

🐙 Kraken Upscale & Tile Calc

Calculate parameters for Ultimate SD Upscale.

Features:

• Connects to Kraken Empty Latent Image via resolution_in and preset_in
• Calculates optimal upscale factor to reach target
• Computes tile width/height for USDU
• Handles seam fix parameters
• Minimum upscale policy (auto, off, 1.5x, 2.0x, custom)
• Generates preview showing predicted vs target resolution

Outputs:

• image_out (IMAGE): Passthrough
• upscale_by (FLOAT): Scale factor for USDU
• tile_width, tile_height (INT): Tile dimensions
• mask_blur, tile_padding (INT): USDU parameters
• seam_fix_width, seam_fix_mask_blur, seam_fix_padding (INT)
• predicted_resolution, target_resolution, target_preset (STRING)

---

🐙 Kraken Preset From Image

Extract dimensions from an image for upscaling calculations.

Features:

• Multiple sizing modes:
  • Factor: Multiply by scale factor (e.g., 2x)
  • Long side: Target longest edge
  • Short side: Target shortest edge
  • Megapixels: Target total megapixels
• Alignment to model-friendly multiples
• Optional max width/height clamping

Outputs:

• preset_out (STRING): Resolution as "WxH" (e.g., "2048x1536")

---

Sampling Nodes

🐙 Kraken KSampler

Smart KSampler wrapper with AMP handling for modern models.

Why this exists: WAN/Flow/FP8 models manage their own precision internally. Wrapping them in an additional autocast context causes crashes like "RuntimeError: Unexpected floating ScalarType in at::autocast::prioritize". This node automatically detects such models and disables outer AMP.

Features:

• AMP Mode (auto/on/off): Auto-detect WAN/Flow/FP8 models
• All standard KSampler controls
• Negative prompt handling (auto/use/ignore)
• Built-in decode option (standard or tiled)
• Compatible with ComfyUI 0.3.x+ and PyTorch 2.x

Inputs:

Input	Type	Description
model	MODEL	Diffusion model
positive/negative	CONDITIONING	Prompts
amp_mode	DROPDOWN	auto, on, or off
decode_switch	DROPDOWN	on/off for VAE decode
tiled_decode	DROPDOWN	on/off for large images
tile_size	INT	Tile size for tiled decode

Outputs:

• LATENT: Sampled latent
• IMAGE: Decoded image (or placeholder if decode off)

---

Image Processing Nodes

🐙 Kraken Image Resize

Comprehensive image resizing with multiple methods.

Features:

• Source: Upload or upstream connection
• Resize modes: longest_side, width, height, fit, fill, stretch
• Interpolation: lanczos, bicubic, bilinear, nearest, area
• Option to prevent upscaling smaller images
• Background color for fill mode

Outputs:

• image (IMAGE): Resized image
• width, height (INT): Final dimensions
• info (STRING): Processing summary

---

🐙 Kraken Image Processor

Versatile pre/post-processing for upscaling pipelines.

Features:

• Denoising: Bilateral, Gaussian, median, non-local means, GPU-accelerated Gaussian
• Contrast Enhancement: Auto-contrast, manual contrast/brightness
• Sharpening: Unsharp mask, edge enhance, custom kernel
• Color Correction: Saturation, gamma
• Film Grain: Adjustable amount and size (post-processing only)
• Alpha channel preservation
• Quality metrics output
• 16-bit processing precision option

Modes:

• pre_process: Prepare image for upscaling
• post_process: Refine upscaled image (gentler settings auto-applied)

---

Video / WAN Nodes

🐙 Kraken WAN Helper

Quick parameter selection for WAN video generation.

Features:

• Performance tiers: 720p Quality or 480p Speed
• Duration slider: 1-10 seconds
• FPS options: 8, 12, 16, 24, 30
• Automatic aspect ratio detection from first frame
• First/last frame preprocessing
• Bookend frame option for looping

Outputs:

• processed_first_frame, processed_last_frame (IMAGE)
• width, height (INT)
• frames(length) (INT): Total frame count
• fps (FLOAT)
• resolution_text (STRING)

---

🐙 Kraken Last Frame + Meta

Extract last frame from video for creating extended sequences.

Use case: Generate 5-second video clips, then use the last frame of each as the first frame of the next to create seamless longer videos.

Features:

• Extracts last frame from image batch
• Forwards all WAN metadata (width, height, frames, fps)
• Optional frame duplication

Outputs:

• last_frame (IMAGE): The last frame
• first_frame_out (IMAGE): Same as last_frame (for chaining)
• width, height, frames(length), fps, resolution_text: Metadata passthrough

---

Typical Workflow Examples

Basic Image Generation with LoRAs

Kraken Checkpoint Loader --> Kraken LoRA Loader (3) --> Kraken KSampler
                                      ^
Kraken Unbound Prompt --> CLIP Text Encode --+

Upscaling with Ultimate SD Upscale

Kraken Empty Latent Image
      |
      +--> resolution_out --> Kraken Upscale & Tile Calc --> Ultimate SD Upscale
      |                              |
      +--> preset_out ---------------+

After USDU:
Ultimate SD Upscale --> Kraken Resolution Helper --> Kraken Image Processor --> Save Image

WAN Video Generation

Kraken WAN Prompt Splitter --> WAN Text Encode
                                     |
Kraken WAN Helper -----------------> WAN Generation --> Kraken Last Frame + Meta
                                                              |
                                          (Loop back to WAN Generation as first_frame)

---

Configuration

CivitAI API Key (for LoRA trigger words)

Create or edit ~/Documents/ComfyUI/user/kraken_config.json:

{
    "civitai_api_key": "your_api_key_here"
}

The key is optional - public model data can be fetched without it.

Ollama Setup (for Prompt Chat)

1. Install Ollama: https://ollama.ai
2. Pull a model: ollama pull llama3.2
3. Configure host in the node (default: http://localhost:11434)

---

Node List Summary

Node	Category	Description
🐙 Kraken Unbound Prompt	Prompt	Vision-enabled prompt builder with styles
🐙 Kraken WAN Prompt Splitter	Prompt	Split & style prompts for WAN video
🐙 Kraken Ollama Prompt Chat	Prompt	LLM-powered prompt enhancement
🐙 Kraken Checkpoint Loader	Loaders	Checkpoint loader with SHA tracking
🐙 Kraken Dual CLIP Loader	Loaders	Dual text encoder for Flux/SD3
🐙 Kraken LoRA Loader (3)	Loaders	3-slot LoRA with CivitAI integration
🐙 Kraken Empty Latent Image	Latent	Latent with aspect ratio presets
🐙 Kraken Resolution Helper	Resolution	Enforce target resolution
🐙 Kraken Upscale & Tile Calc	Resolution	Calculate USDU parameters
🐙 Kraken Preset From Image	Resolution	Extract dimensions from image
🐙 Kraken KSampler	Sampling	Smart sampler with AMP handling
🐙 Kraken Image Resize	Image	Comprehensive image resizing
🐙 Kraken Image Processor	Image	Pre/post-processing pipeline
🐙 Kraken WAN Helper	Video	WAN video parameter helper
🐙 Kraken Last Frame + Meta	Video	Extract last frame for looping

---

License

MIT License - See individual file headers for details.

---

Credits

Created by The Kraken (@KrakenUnbound)

Built for the ComfyUI community.