Stop guessing which Sampler, Scheduler, Prompt, Denoise, Model, Lora or CFG value works best. This custom node suite allows you to generate massive Cartesian product grids, view them in an interactive infinite-canvas dashboard, and refine your settings with a "Revise & Generate" workflow without ever leaving the interface.

📑 Table of Contents

Key Features
Installation
Getting Started
Dashboard Interface
Example Workflows
Preset Configs
Performance Tips
Tips & Best Practices
Troubleshooting
JSON Output Format Reference
API Integration (Config Builder)
File Locations
Keyboard Shortcuts
Changelog
License
Credits

🌟 Key Features

🚀 Powerful Grid Generation

Cartesian Product Engine: Automatically generates every permutation of your input settings. Test unlimited Samplers, Schedulers, CFG scales, Sizes, Prompts, LoRA combinations all in one go.
Visual Config Builder: A point-and-click GUI for building complex sampler configurations — no JSON editing required. Searchable dropdowns, drag sliders, folder expansion, LoRA stacking, and trigger word management all built in.
Non-Standard Model Support: Full support for SD3, Flux, Z-Image, and other non-standard architectures with automatic latent channel detection.
Multi-Model Support: Test multiple checkpoints in a single run by passing an array of model names or folder paths.
Model Folder Expansion: Use "model": "FolderName/" to test all checkpoints in a folder automatically — perfect for comparing different versions or architectures.
CLIP Skip Support: Control which CLIP layer to use for text encoding with the clip_skip parameter — essential for anime models (typically -2) vs realistic models (typically 0).
Intelligent CLIP Encoding: Automatically detects when multiple models are used and handles CLIP encoding correctly per-model to ensure accurate results.
Batch Encoding on Model Switch: When switching between models, all prompts are batch-encoded at once rather than per-generation, resulting in 3-6x faster encoding for multi-model workflows.
Multi-LoRA Stacking: Layer multiple LoRAs with custom strengths using the + separator. Supports folder expansion for testing entire LoRA directories.
LoRA Trigger Word Filtering: Use lora_omit_triggers to exclude specific trigger words from auto-appended LoRA triggers, giving you fine control over prompts.
Random LoRA Selection: Randomly select LoRAs from folders with [count,strength] or [count,strength,random] syntax. Supports both reproducible (seed-based) and truly random selection modes.
Auto LoRA Trigger Words: Automatically fetches and appends LoRA trigger words from CivitAI API using SHA256 hash lookup. Results are cached locally for offline use.
Multi-Seed Generation: Add extra random variations per config with the add_random_seeds_to_gens parameter — perfect for evaluating consistency.
Smart Caching: Intelligently skips model and LoRA reloading when consecutive runs share the same resources, making generation instant for parameter tweaks.
Stop & Resume: Intelligent skip detection — if you stop a generation mid-run, resuming will skip already-generated images and continue where you left off.
Advanced Skip Logic: Uses conditioning tensor hashing to detect prompt changes even when using pre-encoded conditioning from CLIP nodes.
LoRA Compatibility Detection: Automatically detects and skips incompatible LoRAs with detailed error reporting, preventing log spam.
Graceful Interruption: Cancel button now stops ALL remaining jobs (not just current generation) and saves all completed work including pending remote VAE decodes.
VAE Batching: Includes a vae_batch_size input to batch decode images, significantly speeding up large grid runs.
Live Dashboard Updates: Configure flush_batch_every to update the dashboard incrementally (e.g., every 4 images) instead of waiting for the entire batch to complete.
Remote VAE Support: Offload VAE decoding to remote servers (HuggingFace endpoints or local) for 20-30% faster generation and lower VRAM usage.

🎨 Interactive Dashboard (The "IDE")

Infinite Canvas with Pan/Zoom: Google Maps-style navigation with mouse drag, mousewheel zoom, and keyboard shortcuts.
Virtual Scrolling: Ultra-optimized rendering handles thousands of images smoothly by only loading visible items — scroll through 5000+ images without lag.
Mobile Touch Support: Full pinch-to-zoom and pan gestures on mobile devices with optimized touch controls.
Fullscreen Mode: Click the fullscreen button (⛶) to expand the dashboard to fill your entire screen.
Favorites System: Star your best images with a ⭐ button — favorites are collected in a separate gold JSON bar for easy export.
Smart Filtering: Toggle visibility by Model, Sampler, Scheduler, Denoise, or LoRA type. Shift+Click to isolate a single filter for quick A/B testing.
Intelligent Sorting: Instantly sort your grid by Oldest, Newest, or Fastest (generation time). Your preference is saved to localStorage.
Go to Image #: Jump directly to any image number with the "Go to #" input field in the header.
Auto-Load Sessions: Dashboard automatically loads when generation starts — no manual session name entry needed.
Session Management: Save and load previous testing sessions directly from the UI.
Keyboard Navigation: Space to scroll rows, Arrow Keys to pan, +/- to zoom, 0 to reset, F to auto-fit.

⚡ The "Revise & Generate" Workflow

One-Click Revision: Click "REVISE" on any image to open a detail view.
Complete Metadata View: Shows model, seed, prompts (with trigger words), and all generation parameters.
Instant Tweak: Adjust CFG, Steps, or Sampler for just that specific image.
Generate New: A "GENERATE NEW" button queues the new variation immediately without needing to disconnect wires or change the main batch.
Similarity Reel: The revision modal shows a side-scrolling reel of all other images that share the same seed, allowing for perfect A/B comparison.
Multiple Prompts: Use an array to run multiple prompts in one run without re-running your entire workflow: ["picture of a forest", "mountains at night", "masterpiece, painting of dog"]

🧹 Curation & JSON Export

Rejection System: Click the red "✕" on bad generations to hide them.
Triple JSON Bars (Horizontal Layout):
- Green Bar (Left): Automatically groups all accepted configs into a clean, optimized JSON array ready for copy-pasting.
- Gold Bar (Center): Contains all favorited configs — your best-performing settings.
- Red Bar (Right): Collects all rejected configs so you know exactly what settings to avoid.

📦 Installation

Navigate to your ComfyUI custom_nodes directory:
```
cd ComfyUI/custom_nodes/
```

Clone this repository:

git clone https://github.com/YOUR_USERNAME/ComfyUI-Ultimate-Auto-Sampler-Config-Grid-Testing-Suite.git

Restart ComfyUI.

The Config Builder node is included automatically — no separate installation is required.

Requirements:

ComfyUI (latest version recommended)

File Structure:

ComfyUI/custom_nodes/ComfyUI-Ultimate-Auto-Sampler/
├── config_builder_node.py           # Config Builder backend
├── js/conf_builder/
│   ├── conf-builder-main.js         # Entry point & node registration
│   ├── conf-builder-utilities.js    # Data fetching & parsing
│   ├── conf-builder-ui-components.js # Reusable UI elements
│   └── conf-builder-config-management.js # Rendering & state logic
├── trigger_words.py                 # LoRA trigger word handling
├── batch_encoding.py                # CLIP batch encoding with caching
├── manifest_utils.py                # Manifest file management
├── model_loader.py                  # Model/LoRA loading and patching
├── image_generation.py              # Image generation and sampling
└── generation_orchestrator.py       # Main orchestration layer

Getting Started

1. The Nodes

This suite consists of three nodes found under the sampling/testing category:

Ultimate Sampler Grid (Generator): The engine. It handles model loading, grid generation, and saving.
Ultimate Grid Dashboard (Viewer): The display. It renders the HTML output.
Ultimate Config Builder: The visual GUI for building configs without writing JSON.

Basic Setup:

Add the Generator node.
Connect your Checkpoint, CLIP, and VAE (optional, see "Hybrid Inputs" below).
Add the Viewer node.
Connect the dashboard_html output from the Generator to the input of the Viewer.
(Optional) Add the Config Builder node and connect its configs_json output to the Generator's configs_json input.

2. Generator Node Parameters

Core Settings

ckpt_name: Default checkpoint to use (can be overridden by model in JSON or optional input).
positive_text / negative_text: Prompts. Supports arrays: ["prompt 1", "prompt 2"] or JSON arrays.
seed: Base seed. Each config uses this seed unless overridden.
denoise: Denoise strength(s). Supports single value, array, or comma-separated: "1.0" or "0.8, 0.9, 1.0".

Performance Settings

vae_batch_size: How many images to decode at once.
- 4 (default): Good for 8-12GB VRAM
- 12-24: For 24GB+ VRAM
- -1: Decode all images at once (fastest, but risky)
flush_batch_every: Update dashboard every X images (0 = use VAE batch size).
- 0: Wait until VAE batch is full
- 4: Update dashboard every 4 images (recommended for live monitoring)

Advanced Settings

overwrite_existing:
- False (default): Skip already-generated images (resume mode)
- True: Re-generate everything
add_random_seeds_to_gens: Generate X extra random variations per config.
- 0 (default): Only use base seed
- 3: Generate 3 additional random seed variations per config
- Random seeds are deterministic per base seed — changing base seed generates new random variations
lookup_and_append_lora_triggerwords: Automatically fetch and append LoRA trigger words.
- False (default): Use prompts as-is
- True: Calculate SHA256 hash of each LoRA, query CivitAI API for trigger words, cache results locally, and prepend to prompts
- Cache stored in loras_tags.json for offline use
session_name: Folder name where results are saved (ComfyUI/output/benchmarks/{session_name}/).

3. The JSON Configuration

The configs_json widget determines your grid. It accepts an array of objects. All fields support single values or arrays. You can write JSON by hand or use the Config Builder Node to build it visually.

Basic Example:

[
  {
    "sampler": ["euler", "dpmpp_2m"],
    "scheduler": ["normal", "karras"],
    "steps": [20, 30],
    "cfg": [7.0, 8.0]
  }
]

Generates 16 images (2 samplers × 2 schedulers × 2 steps × 2 cfg)

Advanced Features:

Multi-Model Testing

[
  {
    "sampler": "euler",
    "scheduler": "normal", 
    "steps": 20,
    "cfg": 7.0,
    "model": [
      "sd_xl_base_1.0.safetensors",
      "juggernautXL_v9.safetensors",
      "ponyDiffusionV6XL_v6.safetensors"
    ]
  }
]

Tests 3 different models with the same settings

Folder Expansion

[
  {
    "sampler": "euler",
    "scheduler": "normal",
    "steps": 20, 
    "cfg": 7.0,
    "model": "sdxl_models/"
  }
]

Trailing / tests ALL models in the folder.

Multi-LoRA with Stacking

[
  {
    "sampler": "euler",
    "scheduler": "normal",
    "steps": 20,
    "cfg": 7.0,
    "lora": [
      "None",
      "style_lora.safetensors:0.8:1.0",
      "style_lora.safetensors:0.5:1.0 + detail_lora.safetensors:1.0:1.0",
      "loras_folder/"
    ]
  }
]

LoRA format: filename:strength_model:strength_clip
Stack with + separator: lora1:0.8:1.0 + lora2:1.0:1.0

Resolution Grid

// In resolutions_json input:
[
  [1024, 1024],
  [1024, 1536], 
  [1536, 1024]
]

Multiple Prompts

// In positive_text input (as JSON array):
[
  "a beautiful landscape, mountains, sunset",
  "cyberpunk city at night, neon lights",
  "portrait of a warrior, detailed armor"
]

4. Hybrid Inputs (Optional)

The Generator node features built-in widgets for Model Selection and Prompts, but also has Optional Inputs for flexibility:

Standalone Mode: Use the dropdown menu to select a checkpoint and type prompts into the text boxes.
Hybrid Mode: Connect a MODEL, CLIP, VAE, or CONDITIONING wire. The node will automatically ignore the internal widget and use the connected input instead.
Non-Standard Models: For SD3, Flux, Z-Image, and other architectures:
- Connect optional_model, optional_clip, and optional_vae from your model loader
- Connect optional_positive and optional_negative for pre-encoded conditioning
- The node automatically detects latent channel count (4 for SD1.5/SDXL, 16 for SD3/Flux/Z-Image)
- Skip detection uses conditioning tensor hashing to properly detect prompt changes
Latent Input: Connect a LATENT to use img2img or upscaling workflows.
- For SD3/Flux: Use EmptySD3LatentImage instead of EmptyLatentImage
- Latent dimensions are automatically preserved

🔧 Config Builder Node (Visual UI)

The Config Builder node replaces manual JSON editing with an intuitive visual interface. Instead of writing configuration arrays by hand, you can point and click to select models and LoRAs, drag sliders to adjust strengths, expand entire folders, stack multiple LoRAs, manage trigger words, and preview JSON output in real-time.

All configuration data is automatically synchronized with the Ultimate Sampler Grid node.

Basic Setup

Add the Node: Right-click on the canvas → Add Node → sampling/testing → Ultimate Config Builder
Connect to Sampler:
- Connect the configs_json output to the configs_json input of the Ultimate Sampler Grid (Generator) node
- Connect the session_name output to the session_name input of the generator (optional)
Start Building: The Config Builder UI will appear in the node with one config array ready to customize.

Understanding the Interface

┌─────────────────────────────────────────────────┐
│ 📁 Session Management                          │
│ (Session name, Load session, Refresh button)   │
├─────────────────────────────────────────────────┤
│ 💾 Config Management                           │
│ (Save/Load JSON presets, Auto-save toggle)     │
├─────────────────────────────────────────────────┤
│ ⚙️ Config Arrays                               │
│ ┌─────────────────────────────────────────┐   │
│ │ Config 1                                 │   │
│ │ - Samplers, Schedulers, Steps, CFG      │   │
│ │ - Models Section (expandable)            │   │
│ │ - LoRAs Section (expandable)             │   │
│ │ - Omit Triggers Section                  │   │
│ └─────────────────────────────────────────┘   │
│ ┌─────────────────────────────────────────┐   │
│ │ Config 2...                              │   │
│ └─────────────────────────────────────────┘   │
├─────────────────────────────────────────────────┤
│ 📋 JSON Preview                                │
│ (Real-time preview of generated config)        │
└─────────────────────────────────────────────────┘

Session Management

Located at the top of the interface, this section controls the output session name for your tests.

Session Name: Sets the folder name where results are saved (ComfyUI/output/benchmarks/{session_name}/). Enter a descriptive name like "sdxl_quality_test" or "character_lora_comparison". Changes automatically update the connected Sampler Grid node.
Load Session: Load a previously saved session's metadata from a searchable dropdown.
Refresh Models/LoRAs: Reload the list of available models and LoRAs from disk. Use when you've added new files or the dropdowns are showing stale data.

Config Management (Save/Load Presets)

Save and load configuration presets as JSON files for reuse across projects.

Config Name (Filename): Name for the JSON preset file. Stored at ComfyUI/output/benchmarks/configs/{config_name}.json.
Load Saved Config: Searchable dropdown showing all available .json config files. Replaces current config arrays with the loaded preset.
Save Config Now: Manual save button to persist current configuration to disk. Creates a reusable JSON file containing all config arrays, models, LoRAs, trigger settings, and strength locks.
Auto-Save (2s): Toggle to automatically save the config file every 2 seconds. Recommended for long editing sessions.

Config Arrays

Config Arrays are the building blocks of your testing grid. Each array defines a set of parameters that will be multiplied into all possible combinations.

Fast CivitAI Info Lookup

Basic Parameters

Config Name: Label for this config group (purely organizational, not used in generation).
Samplers: Comma-separated list of samplers to test (e.g., "euler, dpmpp_2m, dpmpp_2m_sde").
Schedulers: Comma-separated list of schedulers (e.g., "normal, karras, exponential").
Steps: Comma-separated list of step counts (e.g., "20, 30, 40").
CFG: Comma-separated list of CFG scale values (e.g., "6.0, 7.0, 8.0").

Iteration Count Display

⏱️ Iterations: 144

Shows the total number of images this config array will generate: Formula: Models × LoRAs × Samplers × Schedulers × Steps × CFG
Updates in real-time as you add/remove items.

Config Array Controls

➕ Add Model — Adds a new model slot to the Models section.
➕ Add LoRA — Adds a new LoRA slot to the LoRAs section.
📋 Duplicate Config — Creates an exact copy of this config array (useful for variations).
🗑️ Delete Config — Removes this entire config array.
▼ Collapse/Expand — Toggle visibility of Models and LoRAs sections.

Models Section

Each model card includes:

Model Name (Searchable): Dropdown with fuzzy search through all available checkpoints. Supports folder structure display (e.g., "subfolder/model.safetensors").
Type Selection:
- Single File (default): Test just this specific model.
- Folder (Separate): Expand folder — each model becomes a separate test. Multiplies iteration count by number of models in folder.
- Folder (Combined): Reserved for future use.
🗑️ Remove Model: Deletes this model from the config (at least one model is always required).

LoRAs Section

Each LoRA card includes:

LoRA Name (Searchable): Dropdown with fuzzy search. Supports special syntax — end with / for folder expansion, end with /* for combined folder stack.
Type Selection:
- Single File: Load this specific LoRA.
- Folder (Separate): Test each LoRA in the folder individually with the same strength settings.
- Folder (Combined): Stack ALL LoRAs in the folder into a single load.
Model Strength Slider: Range 0.0–2.0 (default 1.0). Controls how strongly the LoRA affects the model. Number input accepts values beyond slider range.
CLIP Strength Slider: Range 0.0–2.0 (default 1.0). Controls how strongly the LoRA affects text encoding. Independent from model strength.
🔗 Lock Strengths: Toggle to link model and CLIP strengths so both sliders move together.
🔍 Lookup Trigger Words: Opens the Trigger Word Lookup Modal (fetches from CivitAI API).
🚫 Bypass Trigger Fetch: Toggle to prevent automatic trigger word lookup for this LoRA.
🗑️ Remove LoRA: Deletes this LoRA from the config (at least one is always required).

LoRA Trigger Words

The Config Builder integrates with CivitAI's API to automatically fetch and manage trigger words for your LoRAs.

How It Works:

When you select a LoRA, the node calculates its SHA256 hash.
The hash is sent to CivitAI to retrieve metadata.
Results are cached in ComfyUI/output/benchmarks/loras_tags.json.
Cached data is reused on subsequent loads (offline-capable).

Click the 🔍 Lookup Trigger Words button to open the modal:

╔════════════════════════════════════════════╗
║ LoRA Trigger Word Manager                 ║
║                                            ║
║ LoRA: character_style_v2.safetensors      ║
║ Model: Character Style LoRA v2             ║
║ Base Model: SDXL 1.0                       ║
║ Creator: ArtistName                        ║
║                                            ║
║ Tags: character, anime, style              ║
║                                            ║
║ Trigger Words:                             ║
║ ☑️ character_name                          ║
║ ☑️ red_hair                                ║
║ ☑️ blue_eyes                               ║
║ ☐ masterpiece  (excluded)                 ║
║ ☐ best quality (excluded)                 ║
║                                            ║
║ [✅ Save Selections]  [❌ Cancel]          ║
╚════════════════════════════════════════════╝

Checkboxes to select which trigger words to exclude.
Metadata display showing model info before deciding.
Persist selections — saved exclusions apply to all future generations.
CivitAI link to open the model page in a new tab.

Excluded triggers are stored in the config array's lora_omit_triggers array and automatically filtered during generation.

Trigger Word Placement Control

Control exactly where trigger words are inserted in your prompts. For each LoRA with trigger words, you can choose:

None (Default): Trigger words are prepended to the beginning of the positive prompt. Best for character LoRAs and style LoRAs that need early influence.
Positive Start: Explicitly prepend to positive prompt (same as None/default).
Positive End: Append trigger words to the end of the positive prompt. Best for quality tags and detail LoRAs.
Negative Start: Prepend trigger words to the negative prompt. Best for negative embeddings and "anti-style" LoRAs.
Negative End: Append trigger words to the negative prompt. Best for quality-control negative tags.

Set placement by opening the trigger word modal, selecting placement from the dropdown at the bottom, and saving. The placement is stored in lora_triggerwords_append_settings.

LoRA Bypass States

Bypass states allow you to mark specific LoRAs to skip automatic trigger word fetching.

When to use: LoRAs without triggers, offline work, custom workflows where you're manually adding triggers, or speed optimization for known LoRAs.

Toggle the 🚫 Bypass Trigger Fetch checkbox on a LoRA card. The card shows a ⚠️ indicator when bypassed. Bypass state is saved in the lora_bypass_states object.

Strength Locking

Click the 🔗 Lock Strengths button on a LoRA card to link model and CLIP strengths. When locked, both sliders move together maintaining the ratio. For example, if locked at 1.0/0.8 (ratio 1.25), changing model to 1.2 sets CLIP to 0.96.

Useful for maintaining tested ratios and quickly scaling both strengths up/down together.

Folder Expansion (Models & LoRAs)

Folder expansion allows you to test entire directories of models or LoRAs without manually adding each file.

Model Folder Expansion: Select a folder path ending with / and set type to "Folder (Separate)." All .safetensors files in that folder are tested individually with the same settings.

LoRA Folder Expansion — Separate: Path ending with /, type "Folder (Separate)." Each LoRA in the folder is tested individually. All use the same strength settings, and trigger words are fetched for each automatically.

LoRA Folder Expansion — Combined: Path ending with /*, type "Folder (Combined)." ALL LoRAs in the folder are stacked into a single load. Creates one config with multiple LoRAs stacked at the same strength. Trigger words from all LoRAs are combined. Great for quality enhancement stacks, multi-concept combinations, or testing "everything at once" scenarios.

LoRA Stacking

Combine multiple LoRAs into a single generation by adding multiple LoRA slots to a config array.

Manual Stacking: Click ➕ Add LoRA multiple times and configure each independently with different files, strengths, and trigger settings. All selected LoRAs are applied simultaneously.

Generated config format:

{
  "lora": "character_lora.safetensors:1.0:1.0 + style_lora.safetensors:0.8:0.8 + detail_lora.safetensors:0.5:0.5"
}

Mixed Expansion + Manual: You can combine folder expansion with manual stacking — for example, a combined quality folder plus a specific character LoRA stacked together.

🖥️ Dashboard Interface

Header Bar

Model/Prompt Info: Shows current model and prompt metadata.
Go to Image #: Jump directly to any image by entering its number (shown in bottom-left of cards).
Column Count: Set fixed grid columns or leave at 0 for auto-sizing.
Zoom Controls: ⊙ (reset), − (zoom out), + (zoom in).

Session Controls:
- Dashboard auto-loads when connected to sampler and generation starts.
- LOAD to view previous results by session name.
- SAVE to persist current state to disk.
- DELETE to remove session and all images.
Filter Groups: Click colored buttons to toggle visibility:
- Model (Purple): Filter by checkpoint
- Sampler (Cyan): Filter by sampler type
- Scheduler (Blue): Filter by scheduler
- Denoise (Red): Filter by denoise value
- LoRA (Orange): Filter by LoRA configs
- Shift+Click: Isolate single filter (deselect all others)
Sort Button: Cycles between Sort: Oldest (default), Sort: Newest, and Sort: Fastest (by generation time). Preference is saved to localStorage.
Fullscreen Button (⛶): Expand dashboard to fill entire screen.

Mouse:

Left-click or middle-click drag to pan
Scroll wheel to zoom in/out
Right-click on canvas to focus for keyboard controls

Touch (Mobile/Tablet):

Single finger drag to pan
Two finger pinch/spread to zoom
Tap card to reveal buttons

Keyboard:

Space — Scroll down one row
Shift+Space — Scroll up one row
↑↓←→ — Pan in any direction
+/- — Zoom in/out
0 — Reset zoom to 1:1
F — Auto-fit first row to viewport width

Card Overlays

Bottom Left: Index number (#1, #2, etc.) — used for "Go to #" feature.
Bottom Right: Generation time in seconds.
Top Left (on hover): Red ✕ button to reject/hide image.
Top Right (on hover): Gold ⭐ button to favorite image, and green "REVISE" button to open studio view.

JSON Bars (Bottom — Horizontal Layout)

Green Bar (Left — Accepted): Contains optimized JSON of all currently visible images. Click to select all, then copy-paste back into the configs_json widget to refine your batch.
Gold Bar (Center — Favorites): Contains configs of all images you starred with ⭐. Your best-performing settings in one place.
Red Bar (Right — Rejected): Contains the configs of images you deleted with the ✕ button. Know what to avoid.

Clicking REVISE on a card opens the studio view:

Left: Full-resolution preview.
Top Right — Read-Only Info: Model used, seed number, positive prompt (with trigger words if applicable), negative prompt.
Bottom Right — Adjustable Parameters: Sampler, Scheduler, Steps, CFG, Denoise, LoRA.
Bottom: "Related Variants" reel showing other images with the same seed.
GENERATE NEW: Queues the specific config you just edited.

🎯 Example Workflows

Quick Quality Test (40 images)

[
  {
    "sampler": ["dpmpp_2m", "euler"],
    "scheduler": ["karras", "normal"],
    "steps": [20, 30],
    "cfg": [6.0, 7.0, 8.0],
    "denoise": "1.0"
  }
]

Multi-Model Comparison (9 images)

[
  {
    "sampler": "euler",
    "scheduler": "normal",
    "steps": 25,
    "cfg": 7.0,
    "model": [
      "model_v1.safetensors",
      "model_v2.safetensors", 
      "model_v3.safetensors"
    ]
  }
]

Set add_random_seeds_to_gens: 2 to get 3 variations per model (27 total images).

LoRA Stack Testing (24 images)

[
  {
    "sampler": "euler",
    "scheduler": "normal",
    "steps": 25,
    "cfg": 7.0,
    "lora": [
      "None",
      "style.safetensors:0.6:1.0",
      "style.safetensors:0.8:1.0",
      "style.safetensors:1.0:1.0",
      "style.safetensors:0.8:1.0 + detail.safetensors:0.5:1.0",
      "style.safetensors:1.0:1.0 + detail.safetensors:1.0:1.0"
    ]
  }
]

Random LoRA Selection

Randomly select LoRAs from a folder for variety and experimentation:

Basic: "lora": "XL/Styles/[3,0.85]" — Selects 3 random LoRAs at strength 0.85 (seed-based, reproducible).

Truly Random: "lora": "XL/Styles/[3,0.85,random]" — Add random keyword for non-reproducible selection.

Dual Strength: "lora": "XL/Characters/[2,0.8,0.6]" — Model strength 0.8, CLIP strength 0.6.

Combined with Regular LoRAs:

"lora": "XL/base.safetensors:1.0 + XL/Styles/[2,0.7] + XL/Details/[1,0.5,random]"

Loads base.safetensors always, picks 2 reproducible style LoRAs, and picks 1 truly random detail LoRA.

Full Model Folder Test

[
  {
    "sampler": "dpmpp_2m",
    "scheduler": "karras",
    "steps": 25,
    "cfg": 7.0,
    "model": "realistic_models/"
  }
]

CLIP Skip for Anime Models

[
  {
    "sampler": "euler",
    "scheduler": "normal",
    "steps": 28,
    "cfg": 7.0,
    "model": "anime_model.safetensors",
    "clip_skip": -2
  }
]

clip_skip: 0 (default) — Use last layer (best for realistic models)
clip_skip: -2 — Skip 2 layers (best for anime/illustration models)

Test multiple values: "clip_skip": [0, -1, -2, -3]

LoRA Trigger Word Filtering

[
  {
    "sampler": "euler",
    "steps": 28,
    "cfg": 7.0,
    "lora": "style2.safetensors:0.8:0.6",
    "lora_omit_triggers": ["style", "makeup", "jewelry"]
  }
]

Filters apply to ALL LoRAs in the stack.

Model Folder Expansion

[
  {
    "model": ["SD1.5/", "SDXL/", "Flux/"],
    "sampler": "euler",
    "steps": 28
  }
]

Tests all models from all three folders with the same settings.

LoRA Strength Sweep

Find the optimal strength for a new LoRA using the Config Builder: add the same LoRA 6 times with strengths from 0.4 to 1.5, keep everything else minimal (1 sampler, 1 scheduler), and compare the 6 images side by side.

Quality Enhancement Stack

Create a consistent quality stack using the Config Builder: add 3 LoRAs manually (e.g., detail_enhancer at 1.0, color_boost at 0.6, sharpness at 0.5), use the trigger word modal to set all placements to Positive End, exclude generic tags like "masterpiece" and "best quality", then save as a reusable preset.

📋 Preset Configs

🏆 Group 1: The "Gold Standards" (Reliable Realism)

Tests the 5 most reliable industry-standard combinations.
5 samplers × 2 schedulers × 2 step settings × 2 cfgs = 40 images

[
  {
    "sampler": ["dpmpp_2m", "dpmpp_2m_sde", "euler", "uni_pc", "heun"],
    "scheduler": ["karras", "normal"],
    "steps": [25, 30],
    "cfg": [6.0, 7.0],
    "lora": "None"
  }
]

🎨 Group 2: Artistic & Painterly

Tests creative/soft combinations best for illustration and anime.
2 samplers × 3 schedulers × 3 step settings × 2 cfgs = 36 images

[
  {
    "sampler": ["euler", "dpmpp_2m"],
    "scheduler": ["simple", "beta", "normal"],
    "steps": [20, 25, 30],
    "cfg": [1.0, 4.5],
    "lora": "None"
  }
]

⚡ Group 3: Speed / Turbo / LCM

Tests ultra-fast configs. (Note: Ensure you are using a Turbo/LCM capable model or LoRA).
4 samplers × 3 schedulers × 4 step settings × 2 cfgs = 96 images

[
  {
    "sampler": ["lcm", "euler", "dpmpp_sde", "euler_ancestral"],
    "scheduler": ["simple", "sgm_uniform", "karras"],
    "steps": [4, 5, 6, 8],
    "cfg": [1.0, 1.5],
    "lora": "None"
  }
]

🦾 Group 4: Flux & SD3 Specials

Tests configs specifically tuned for newer Rectified Flow models like Flux and SD3.
2 samplers × 3 schedulers × 3 step settings × 2 cfgs = 36 images

[
  {
    "sampler": ["euler", "dpmpp_2m"],
    "scheduler": ["simple", "beta", "normal"],
    "steps": [20, 25, 30],
    "cfg": [1.0, 4.5],
    "lora": "None"
  }
]

🧪 Group 5: Experimental & Unique

Tests niche combinations for discovering unique textures.
6 samplers × 4 schedulers × 5 step settings × 4 cfgs = 480 images

[
  {
    "sampler": ["dpmpp_3m_sde", "ddim", "ipndm", "heunpp2", "dpm_2_ancestral", "euler"],
    "scheduler": ["exponential", "normal", "karras", "beta"],
    "steps": [25, 30, 35, 40, 50],
    "cfg": [4.5, 6.0, 7.0, 8.0],
    "lora": "None"
  }
]

🔧 Performance Tips

For Large Batches (1000+ images)

Set flush_batch_every: 10-20 to see progress updates without overwhelming the browser.
Use vae_batch_size: 8-12 (balance between speed and VRAM).
Enable overwrite_existing: False so you can stop/resume safely.

For Multi-Model Testing

Use identical LoRA/prompt settings across models for fair comparison.
Use add_random_seeds_to_gens: 2-3 to evaluate model consistency.

For Memory-Constrained Systems

Lower vae_batch_size to 1-2.

💡 Tips & Best Practices

Organization

Use descriptive config names: "SDXL_Quality_Test_v2" instead of "config1".
Group related tests in separate config arrays.
Save presets early using auto-save to prevent data loss.
Version your configs with numbers in saved config names.

Config Builder Workflow Efficiency

Duplicate instead of recreate: Use the Duplicate Config button for variations.
Load templates: Keep a library of saved configs for common test scenarios.
Leverage the trigger word modal: Bulk-exclude unwanted triggers once — they persist.
Lock strengths for ratios: When you find good ratios (e.g., 1.0/0.8), lock them.
Collapse unused sections: Minimize Models/LoRAs sections when not actively editing.

Trigger Word Management

Exclude generic terms like "masterpiece", "best quality", "highres" — they often appear but may be redundant.
Use placement strategically: character/style triggers → Positive Start, quality/detail triggers → Positive End, anti-style triggers → Negative Start.
Bypass technical LoRAs — technique LoRAs (e.g., "detail tweaker") rarely have meaningful triggers.

Testing Strategy

Start small: 1 model × 1 LoRA × 2 samplers = 2 images.
Expand gradually after confirming the basics work.
Use separate config arrays — don't mix unrelated tests in one array.
Monitor iteration counts — the ⏱️ display updates in real-time. Watch for unexpected explosions.
Save before large grids — always save your config before running 500+ image generations.

⚠️ Troubleshooting

Please check the GitHub Issues list and post any issues there after searching. (Pro-tip: delete state:open from the search to see closed issues too.)

Generation Issues

"Session not found": Ensure the session_name matches a folder inside ComfyUI/output/benchmarks/.
OOM Errors: Lower the vae_batch_size to 1 or 2.
Images not resuming: Make sure overwrite_existing: False. Check console for skip messages.
Random seeds different each run: Intentional — random seeds are tied to the base seed.
"mat1 and mat2 shapes cannot be multiplied": Model architecture mismatch. For SD3/Flux/Z-Image, ensure you connect ALL optional inputs. Check LoRA compatibility.
Dashboard not auto-loading: Ensure the dashboard node is connected to the sampler's dashboard_html output. Try the "RELOAD / SHOW SESSION" button.

Dashboard Issues

Cards not appearing: Click inside the viewport area first to give it focus.
Can't scroll/pan: Right-click on the canvas area to focus it, or left-click drag.
Slow performance: Virtual scrolling handles 5000+ images. If slow, close other tabs, reduce browser zoom to 100%, or clear localStorage.
Images not loading: Scroll slower to give the lazy loader time to fetch.

LoRA & Trigger Word Issues

Trigger words not appearing: Enable lookup_and_append_lora_triggerwords in the sampler node. Ensure internet connection for first-time lookup. Check loras_tags.json for cached results.
"INCOMPATIBLE LORA DETECTED": Normal — the node automatically skips incompatible LoRAs. Ensure your LoRAs match your model architecture.
Random LoRA selection not working: Ensure folder path includes / before the brackets (e.g., "XL/Styles/[3,0.85]"). Check console for debug messages.
Trigger words not filtered with lora_omit_triggers: Ensure lookup_and_append_lora_triggerwords is enabled. Filtering is case-insensitive and handles trailing commas.

Model & CLIP Issues

Images look wrong with multiple models: System handles this automatically. You should see ⚠️ Multiple models detected - pre-encoding DISABLED in console.
CLIP skip not affecting results: Verify the model supports it. Use 0 for realistic, -2 for anime. SDXL may be less sensitive than SD1.5.
"No files found in folder": Ensure trailing / in path. Folder names are case-sensitive on Linux/Mac.

Config Builder UI Issues

Dropdowns not showing models/LoRAs: Click the 🔄 Refresh Models/LoRAs button.
UI not appearing or showing blank: Check browser console (F12) for JavaScript errors. Ensure the node is "Ultimate Config Builder." Refresh the page.
Changes not saving: Ensure auto-save is enabled OR manually click "Save Config Now." Check that the config name field is filled. Verify write permissions to ComfyUI/output/benchmarks/configs/.
"No triggers found" for a known LoRA: LoRA may not be on CivitAI or hash mismatch. Manually add triggers or enable Bypass Trigger Fetch.
Trigger words not updating after exclusions: Open the modal again and verify checkboxes. Click "Save Selections" (not just closing the modal). Check JSON preview for lora_omit_triggers.
API timeout or slow lookups: First lookup requires hashing + API call (10-30 seconds). Subsequent lookups use cache (instant). Bypass trigger fetch for offline work.
Iteration count unexpectedly high: Check for folder expansion in Models/LoRAs sections and accidental commas in CSV fields.
Config not loading from saved file: Verify file exists in ComfyUI/output/benchmarks/configs/. Check JSON syntax. Try manually editing and re-saving.
LoRA stacking not working: Check JSON preview — LoRAs should join with +. Verify each has format "name:model_str:clip_str". Don't mix "None" with actual LoRAs.
Sampler Grid not receiving configs: Verify the configs_json wire is connected. Check JSON preview. Try disconnecting and reconnecting. Queue a prompt to force update.

Remote VAE Issues

Not working: Check endpoint is set correctly. Verify server is running. Look for 🌐 Remote VAE worker started in console.
Hangs on "Waiting for remote VAE": Server may have crashed. Check server logs. Restart and try again.

Interrupt/Cancel Issues

Cancel doesn't stop generation: Ensure latest version with interrupt handling. Should see 🛑 INTERRUPTED - Stopping all jobs in console.
Work lost after canceling: Manifest is saved on interrupt. Check ComfyUI/output/benchmarks/{session_name}/manifest.json.

Browser Compatibility

Chrome/Edge: Full support ✅
Firefox: Full support ✅
Safari: Mostly works, some keyboard shortcuts may conflict
Mobile: Full touch support ✅ (iOS Safari, Android Chrome tested)

JSON Output Format Reference

The Config Builder generates JSON in the exact format required by the Ultimate Sampler Grid:

[
  {
    "sampler": ["euler", "dpmpp_2m"],
    "scheduler": "normal",
    "steps": 25,
    "cfg": [6.0, 7.0, 8.0],
    "lora": "style.safetensors:0.8:0.8 + detail.safetensors:0.5:0.5",
    "model": "base_xl.safetensors",
    "lora_omit_triggers": ["masterpiece", "best quality"],
    "lora_triggerwords_append_settings": {
      "style.safetensors": "positive_start",
      "detail.safetensors": "positive_end"
    },
    "lora_bypass_states": {
      "technique_lora.safetensors": true
    },
    "lora_strength_lock": {
      "style.safetensors": true
    }
  }
]

Field Types:

Single value: "euler" or 7.0
Array: ["euler", "dpmpp_2m"] or [6.0, 7.0, 8.0]
LoRA string: "name:model_str:clip_str" — stack with +
Folder: "path/" (separate) or "path/*" (combined)

API Integration (Config Builder)

For advanced users, the Config Builder exposes several backend endpoints:

`/configbuilder/lookup_triggers` (POST)

Bulk lookup trigger words for multiple LoRAs.

Request:

{
  "loras": ["lora1.safetensors", "lora2.safetensors"]
}

Response:

{
  "triggers": {
    "lora1.safetensors": ["trigger1", "trigger2"],
    "lora2.safetensors": ["trigger3"]
  }
}

`/configbuilder/lookup_lora_metadata` (POST)

Fetch complete metadata for a specific LoRA from CivitAI.

Request:

{
  "lora_name": "character_lora.safetensors"
}

Response:

{
  "metadata": {
    "name": "Character LoRA v2",
    "model_name": "Character Series",
    "trained_words": ["character", "outfit"],
    "base_model": "SDXL 1.0",
    "description": "...",
    "tags": ["character", "anime"],
    "url": "https://civitai.com/models/12345",
    "hash": "abc123...",
    "creator": "ArtistName"
  }
}

`/configbuilder/refresh_models` (POST)

Signal to clear frontend caches (called by ComfyUI when refreshing node definitions).

File Locations

Generated Files

Config Presets:

ComfyUI/output/benchmarks/configs/
├── my_quality_test.json
├── model_comparison.json
└── lora_strength_sweep.json

Session Output:

ComfyUI/output/benchmarks/{session_name}/
├── manifest.json
├── image_001.png
├── image_002.png
└── ...

LoRA Trigger Cache:

ComfyUI/output/benchmarks/loras_tags.json

LoRA Metadata:

ComfyUI/output/benchmarks/model-data/
├── lora_name/
│   └── metadata.json

Keyboard Shortcuts

Dashboard

Shortcut	Action
`Space`	Pan down one row
`Shift+Space`	Pan up one row
`↑↓←→`	Pan in any direction
`+/-`	Zoom in/out
`0`	Reset zoom to 1:1
`F`	Auto-fit first row to viewport width

Config Builder

Shortcut	Action
`Ctrl+F / Cmd+F`	Search within searchable dropdowns (when focused)
`Enter`	Confirm selection in dropdown
`Escape`	Close dropdown without selecting
`Tab`	Navigate between input fields

📝 Changelog

Update 2/5/26 — Code Refactoring & Performance Improvements

🏗️ Major Code Refactoring: Reorganized codebase into 6 modular files for better maintainability (trigger_words.py, batch_encoding.py, manifest_utils.py, model_loader.py, image_generation.py, generation_orchestrator.py).
🎯 CLIP Skip Support: Control CLIP layer usage with clip_skip parameter. Essential for anime models (-2) vs realistic models (0). Supports arrays for testing multiple values.
🧠 Intelligent CLIP Encoding: Multi-model workflows now handle CLIP correctly with automatic detection and per-model encoding.
⚡ Batch Encoding on Model Switch: 3-6x faster encoding for multi-model workflows.
🗑️ LoRA Trigger Word Filtering: New lora_omit_triggers parameter to exclude specific trigger words.
📁 Model Folder Expansion: Use "model": "FolderName/" to test all checkpoints in a folder.
🛑 Graceful Interruption: Cancel now stops ALL jobs, flushes pending batches, waits for remote VAE, and saves manifest.
🌐 Remote VAE Fixes: Fixed initialization, job queuing, and interrupt handling.
🔧 str_model/str_clip Removal: Deprecated config fields removed. Strengths now specified in LoRA string.

Update 2/5/26 — Random LoRA Selection

🎲 Random LoRA Selection: [count,strength] syntax for seed-based selection, [count,strength,random] for truly random, dual strength support, cross-platform paths, and full trigger word integration.

Update 1/14/26 — Major Feature Update

🎯 Non-Standard Model Support: SD3, Flux, Z-Image with automatic latent channel detection.
⭐ Favorites System: Star images with dedicated gold JSON bar.
🎨 Horizontal JSON Bars: Redesigned three-bar layout.
🔍 LoRA Auto Trigger Words: CivitAI API integration with local caching.
🚫 LoRA Compatibility Detection: Automatic skip with clear error messages.
⌨️ Shift+Click Filter Isolation.
🎯 Go to Image #.
🚀 Dashboard Auto-Load.
📱 Mobile Touch Support.
🔐 Conditioning Change Detection.
📋 Enhanced Revise Modal.

Update 1/11/26 — Major Overhaul

✨ Virtual Scrolling: Handles 5000+ images.
🖼️ Fullscreen Mode.
🔄 Multi-Model Support with folder expansion.
🎨 Multi-LoRA Stacking with + separator.
🎲 Multi-Seed Generation.
⏸️ Stop & Resume.
⌨️ Keyboard Navigation.
📊 Live Updates with flush_batch_every.
💾 Persistent Settings.
🎯 Auto-Fit Zoom.
⚡ Massive performance refactoring.

📜 License

MIT License. Feel free to use, modify, and distribute.

🙏 Credits

Created for the ComfyUI community. Special thanks to all contributors and testers who helped refine this tool.

Special Features Powered By:

CivitAI API for trigger word lookups
ComfyUI's object_info system for model/LoRA discovery

Star this repo if you find it useful! ⭐

Support development: Buy me a coffee on Ko-fi ☕

Version: 2.0
Last Updated: February 2026

README.md Unescape Escape

ComfyUI Ultimate Auto Sampler Config Grid Testing Suite

📑 Table of Contents

🌟 Key Features

🚀 Powerful Grid Generation

🎨 Interactive Dashboard (The "IDE")

⚡ The "Revise & Generate" Workflow

🧹 Curation & JSON Export

📦 Installation

Getting Started

1. The Nodes

2. Generator Node Parameters

Core Settings

Performance Settings

Advanced Settings

3. The JSON Configuration

Multi-Model Testing

Folder Expansion

Multi-LoRA with Stacking

Resolution Grid

Multiple Prompts

4. Hybrid Inputs (Optional)

🔧 Config Builder Node (Visual UI)

Basic Setup

Understanding the Interface

Session Management

Config Management (Save/Load Presets)

Config Arrays

Fast CivitAI Info Lookup

Basic Parameters

Iteration Count Display

Config Array Controls

Models Section

LoRAs Section

LoRA Trigger Words

Trigger Word Lookup Modal

Trigger Word Placement Control

LoRA Bypass States

Strength Locking

Folder Expansion (Models & LoRAs)

LoRA Stacking

🖥️ Dashboard Interface

Header Bar

Toolbar

Navigation & Controls

Card Overlays

JSON Bars (Bottom — Horizontal Layout)

Revision Modal

🎯 Example Workflows

Quick Quality Test (40 images)

Multi-Model Comparison (9 images)

LoRA Stack Testing (24 images)

Random LoRA Selection

Full Model Folder Test

CLIP Skip for Anime Models

LoRA Trigger Word Filtering

Model Folder Expansion

LoRA Strength Sweep

Quality Enhancement Stack

📋 Preset Configs

🏆 Group 1: The "Gold Standards" (Reliable Realism)

🎨 Group 2: Artistic & Painterly

⚡ Group 3: Speed / Turbo / LCM

🦾 Group 4: Flux & SD3 Specials

🧪 Group 5: Experimental & Unique

🔧 Performance Tips

For Large Batches (1000+ images)

For Multi-Model Testing

For Memory-Constrained Systems

💡 Tips & Best Practices

Organization

Config Builder Workflow Efficiency

Trigger Word Management

Testing Strategy

⚠️ Troubleshooting

Generation Issues

Dashboard Issues

LoRA & Trigger Word Issues

Model & CLIP Issues

Config Builder UI Issues

README.md

`/configbuilder/lookup_triggers` (POST)

`/configbuilder/lookup_lora_metadata` (POST)

`/configbuilder/refresh_models` (POST)