Eric's Prompt Enhancers for ComfyUI

A comprehensive suite of 18 nodes for ComfyUI using local LLMs (LM Studio or Ollama). Transform simple prompts into detailed, platform-optimized descriptions for video and image generation.

---

📦 Quick Start

cd ComfyUI/custom_nodes/
git clone https://github.com/EricRollei/Local_LLM_Prompt_Enhancer.git video_prompter
cd video_prompter
pip install -r requirements.txt

Restart ComfyUI. All nodes will appear under: Add Node → Eric Prompt Enhancers

---

🆕 New in 2.0

Version 2.0 keeps every existing node and every saved workflow, and adds three things underneath them. The documents linked here are current.

A real wildcard engine. The Dynamic Prompts / Impact Pack grammar (__name__, {a|b|c}, weights, ranges, variables, tag queries such as __objects[metal AND handheld]__), read straight from ComfyUI/models/wildcards, in every enhancer node. Picks are handed to the LLM as concepts to express, as exact strings that must survive (__=lora_triggers__), or as loose inspiration (__?moods__). New nodes: Wildcard Processor, Semantic Variation (moves picks to semantic neighbours, with named direction axes) and Concept Survival Check. See docs/WILDCARD_GUIDE.md.

Platform profiles. Everything the nodes know about a target model is one YAML file in profiles/: format, length, negative prompt or none, quality tokens or none, how text and dialogue are written, and - where the maker publishes one - the maker's own rewriter prompt, downloaded once and cached. Over 60 models, open and hosted: FLUX.2, Krea 2, Z-Image, Qwen-Image, HunyuanImage 3, Ideogram 4, Illustrious, NoobAI; MiniMax H3, Wan 2.2, LTX-2, HunyuanVideo 1.5; and for API use Seedream 5, Nano Banana, GPT Image, Midjourney V8, Seedance 2.5, Kling 3, Veo 3.1, Wan 3.0 and others. Your own go in profiles_user/. The new Platform Prompt Rewriter node writes for any of them, including the structured formats (MiniMax H3 fields, FLUX.2 and Ideogram JSON, edit instructions). See PLATFORM_NOTES.md.

Prompts stay inside the text-encoder limit. A text encoder silently drops whatever runs past its limit (Krea 2: 512 tokens) - from the END, which is where LoRA trigger words usually go. The Platform Prompt Rewriter now counts the answer the way the model will (the Qwen tokenizer for Qwen-encoder models, from files already on disk; a conservative estimate otherwise) and aims below the limit with headroom for triggers. length_guard (appended): retry+trim asks the LLM once for a tighter version, then trims; trim trims only; off reports only. Trimming drops sentences just before the last one, so the subject (first) and medium / style (last) survive, and verbatim wildcard picks are never cut. The report shows length: N/512 tokens.

Surreal as a procedure. The surreal preset and genre now run an ordinary-scene inventory, generator, critic and structured compiler grounded in how Magritte, Ernst, Dali, Varo, Uelsmann, Atget and Man Ray actually worked: one anomaly stated as plain fact against a credible world, chosen for affinity rather than distance. The compiler removes replaced objects before the platform rewriter and requires visible physical proof of the anomaly. The Surreal Premise node shows every candidate, the structured plan and the reasoning. See docs/SURREAL_GUIDE.md.

The video nodes have a target_model. The three video expanders were written around Wan 2.2. They now write for any video profile, and the default is MiniMax H3. Choose wan22_video to get the original behaviour exactly. Note for existing workflows: a workflow saved before 2.0 has no value for this new input, so it loads with the default (MiniMax H3) - set it to wan22_video once and save.

---

🎯 The Nodes

All 18 nodes live under the "Eric Prompt Enhancers" category in ComfyUI. The five LLM expanders come first, then platform output, wildcards and variation, the prompt builders, and the prompt database.

1. 🎬 Video Prompt Expander

Simple video prompt expansion with style presets.

• 5 Expansion Tiers: auto, basic, enhanced, advanced, cinematic (auto detects the right tier)
• 6 Style Presets: Cinematic, Surreal, Action, Stylized, Noir, Random
• Auto-Variation: Generate up to 3 unique variations
• Best for: Quick video prompt expansion from short ideas

2. 🎬 Video Prompt Expander (Advanced) ⭐ NEW v1.8

Granular control over video aesthetics with 50+ detailed settings.

NEW in v1.8:

• 4 Operation Modes:
  • expand_from_idea - Expand short concepts
  • refine_existing - Polish existing prompts
  • modify_style - Change aesthetic while keeping subject
  • add_details - Enrich existing prompts
• Clear Detail Levels: Concise, Moderate, Detailed, Exhaustive (with tooltips!)
• Optional Image Input: Analyze images with Qwen3-VL for image-to-video workflows
• Auto Mode Detection: Automatically switches to image-to-video when image provided

Controls:

• Lighting: Light source (10 options), Lighting type (13 options), Time of day (9 options)
• Camera: Shot size (10 options), Composition (8 options), Lens (6 options), Angle (10 options), Movement (17 options)
• Visual: Color tone (7 options), Visual style (15 options), Visual effects (11 options)
• Character: Emotion (13 options)

Best for: Professional video generation with precise aesthetic control

3. 🖼️➡️🎬 Image-to-Video Prompt Expander

Vision model analyzes images and adds motion descriptions for video generation.

• Automatic Scene Understanding: Qwen3-VL vision analysis
• Motion Description Generation: AI-generated movement and action
• Style Integration: Combines visual analysis with expansion tiers
• Best for: Converting static images into video prompts

4. 🖼️➡️🖼️ Image-to-Image Prompt Expander

Platform-aware image-to-image prompt generation.

• 6 Platforms: Flux, SDXL, Wan 2.2, HunyuanImage, Qwen-Image, Qwen-Image-Edit
• Transformation Controls: Style transfer, detail level, creativity settings
• Vision Analysis: Understands source image characteristics
• Best for: Image transformation and editing workflows

5. 📝➡️🖼️ Text-to-Image Prompt Enhancer (v1.7)

Advanced multi-platform image prompt enhancement with extensive creative controls.

• 8 Platforms: Flux, SDXL, Pony Diffusion, Illustrious XL, Chroma, Qwen Image, Qwen Edit, Wan Image
• Reference Images: Optional 1-2 image inputs with visual analysis
• Genre Styles: 22 styles (cinematic, horror, cyberpunk, steampunk, noir, fantasy, etc.)
• Prompt Length: 6 options (very_short to very_long, 20-400 tokens)
• Subject Controls: Framing (14 options), Pose (17 options)
• Advanced Settings: Camera angle, composition, lighting (source/quality), weather, time, color mood
• Special Syntax: Emphasis (keyword:1.5) and Alternation {a|b|c} support
• Best for: Professional image generation with platform-specific optimization

6. 🎞️ Alice T2V Prompt Refiner

Refines rough text and/or reference images into structured Alice text-to-video prompts.

• Text-only or image-guided: optional Qwen3-VL vision analysis of reference images
• Best for: Alice / text-to-video prompts from a loose idea

— Platform output —

7. ✍️ Platform Prompt Rewriter

Rewrites a finished prompt for any platform profile (60+ open and hosted models), in that model's exact format, and keeps it inside the text encoder's token limit so trigger words aren't silently dropped.

• 60+ profiles: FLUX.2, Krea 2, Qwen-Image, Ideogram 4, Illustrious, MiniMax H3, Wan 2.2, Veo 3.1, and more (profiles/; your own go in profiles_user/)
• Structured formats: MiniMax H3 fields, FLUX.2 / Ideogram JSON, edit instructions
• length_guard: counts tokens the way the model will and trims to fit, never cutting verbatim wildcard picks
• Outputs: prompt, negative_prompt, report, system_prompt

8. 🪄 Ideogram 4 Magic Prompt

Builds an Ideogram 4 "magic prompt" (structured JSON plus a high-level line) with a recommended canvas size.

• Outputs: prompt_json, high_level, width, height, report
• Best for: Ideogram 4 generation

— Wildcards & variation —

9. 🎲 Wildcard Processor

Resolves wildcards, variants, variables and tag queries in a prompt — the Dynamic Prompts / Impact Pack grammar (__name__, {a|b|c}, weights, ranges, variables, __objects[metal AND handheld]__), read from models/wildcards.

10. 🔀 Semantic Variation

Varies a wildcard prompt by moving each random pick to a semantic neighbour, along named direction axes.

11. ✅ Concept Survival Check

Measures whether the key concepts (a surreal anomaly, exact LoRA triggers, ...) survived the LLM rewrite.

• Outputs: report, all_survived, missing, carried_by_json

12. 🌀 Surreal Premise

Turns an ordinary scene into one anomaly stated as plain physical fact, with an instruction block that protects it through the rewrite, a Magritte-style title and counter-terms. Eight generator modes feed one critic.

• Its llm_instructions plug into any enhancer's wildcard_instructions
• See docs/SURREAL_GUIDE.md

— Prompt builders (click, don't type) —

13. 🧩 Wildcard Prompt Composer

Builds a whole prompt section by section with clickable pills instead of typed wildcard syntax. Holds every section (Format, Subject, Action, Setting, Light & atmosphere, Composition, Camera, Style & finish, plus optional Text) in the order the chosen platform wants. Lock a section to hold it while the seed re-rolls, drag to reorder, save as a template.

14. 👤 Wildcard Character Composer

The same node limited to the Subject section, for a character you reuse across several prompts or branches.

• See docs/PROMPT_BUILDER_GUIDE.md

— Prompt database —

15. 💾 Save Prompt to DB

Saves a prompt (and optional negative) to the local SQLite database, passing the prompt through unchanged so the node can sit inline before a sampler.

16. 🔎 Search Prompt DB

Searches the database by keyword, category, platform, rating, tags or favourites, and returns the top match.

17. 📥 Import Prompts to DB

Bulk-imports prompts from external files (.txt, .csv, and more) into the local database.

18. 🎰 Browse / Random Prompt DB

Fetches a specific prompt by ID, or a random/filtered prompt by seed.

---

✨ Key Features

🤖 Local LLM Support

• LM Studio: OpenAI-compatible API (recommended)
• Ollama: Simple CLI-based LLM server
• Qwen3-VL (Optional): Local vision model for image analysis
• Privacy: All processing happens locally, no data sent to cloud

🎨 Platform-Specific Optimization

Each platform has unique requirements and prompting styles:

Platform	Style	Optimal Length	Specialization
Flux	Natural language	75-150 tokens	Photography, artistic
SDXL	Natural/tags hybrid	40-75 tokens	Versatile, balanced
Pony Diffusion	Booru tags	Tag count	Anime, characters
Illustrious XL	Danbooru tags	Tag count	Detailed anime
Chroma/Meissonic	Detailed natural	100-200 tokens	Complex scenes
Qwen Image	Technical descriptions	Medium	General purpose
Qwen Edit	Edit instructions	Medium	Image editing
Wan Image	Cinematography	60-120 tokens	Professional video stills

🔧 Special Syntax (NEW v1.6.1)

Emphasis (Weight Control)

(keyword:1.5)    # Increase importance 1.5x
(keyword:0.5)    # Decrease importance 0.5x
(red hair:2.0)   # Double the weight

Alternation (Random Selection)

{cat|dog|rabbit}                    # Picks one randomly
{red|blue|green} dress              # Random color
{elegant|casual} woman with (detailed face:1.5)

Combined

A {tall|short} woman with (dark hair:1.4) wearing a {red|blue|green} (dress:1.2)

🎯 Intelligent Features

• Auto-Detection: Analyzes input complexity and selects optimal detail level (Video nodes)
• Multiple Variations: Generate up to 3 unique variations in one run
• Keyword Integration: Automatically include LoRA triggers and custom terms
• Smart Negatives: Platform-optimized negative prompts
• File Export: Save prompts with complete metadata
• Wildcard Support: Random element selection for variety

---

🚀 Installation

Prerequisites

• ComfyUI: Installed and working
• Python 3.8+: Usually included with ComfyUI
• LLM Backend: LM Studio or Ollama (see setup below)

Step 1: Install the Node

Method 1: Git Clone (Recommended)

cd ComfyUI/custom_nodes/
git clone https://github.com/EricRollei/Local_LLM_Prompt_Enhancer.git video_prompter
cd video_prompter
pip install -r requirements.txt

Method 2: Manual Install

1. Download this repository as ZIP
2. Extract to ComfyUI/custom_nodes/video_prompter/
3. Install dependencies:

   cd ComfyUI/custom_nodes/video_prompter/
   pip install -r requirements.txt
   

Step 2: Setup LLM Backend

Choose one of these options:

Option 1: LM Studio (Recommended)

1. Download from: https://lmstudio.ai/
2. Load a model (recommended: Llama 3 8B or similar)
3. Start the server (Settings → Server → Start Server)
4. Default endpoint: http://localhost:1234/v1

Option 2: Ollama

1. Install from: https://ollama.ai/
2. Pull a model: ollama pull llama3
3. Server runs automatically at: http://localhost:11434

Step 3: Restart ComfyUI

All nodes will appear under: Add Node → Eric Prompt Enhancers

Optional: Qwen3-VL Vision Model

For image analysis in Image-to-Video and Image-to-Image nodes:

pip install transformers>=4.42.0 accelerate>=0.30.0 huggingface_hub>=0.23.0 bitsandbytes>=0.43.0

See VISION_BACKEND_GUIDE.md for details.

---

📖 Usage Guide

Basic Workflow (Video Prompt Expander)

1. Add Node: Right-click → Eric Prompt Enhancers → Video Prompt Expander
2. Enter Prompt: "A cat playing piano in a cozy room"
3. Select Preset: Choose "cinematic" for film-like quality
4. Choose Detail Level: "detailed" for ~400-500 words
5. Configure LLM:
   • Backend: lm_studio
   • Model: llama3
   • Endpoint: http://localhost:1234/v1
6. Generate: Run the workflow
7. Use Output: Connect positive_prompt_1 to your video generator

Advanced Workflow (Advanced Node)

1. Add Node: Video Prompt Expander (Advanced)
2. Choose Operation Mode:
   • expand_from_idea: For short concepts
   • refine_existing: For polishing existing prompts
   • modify_style: To change aesthetic
   • add_details: To enrich prompts
3. Select Detail Level: Choose based on desired output length
4. Set Aesthetic Controls: Configure camera, lighting, style, etc.
5. Optional: Add reference image for image-to-video mode
6. Generate: Prompts incorporate your specific controls

Image-to-Video Workflow

1. Add Node: Image-to-Video Prompt Expander (or Advanced with image)
2. Connect Image: Link your image to the reference_image input
3. Add Direction: "Camera slowly zooms in on the subject"
4. Generate: Vision model analyzes image and incorporates it into prompt

Text-to-Image Workflow

1. Add Node: Text-to-Image Prompt Enhancer
2. Select Platform: Choose your target platform (Flux, SDXL, etc.)
3. Set Length: Choose prompt length (medium recommended)
4. Configure Controls: Set genre, framing, lighting, etc.
5. Use Syntax: Add emphasis (keyword:1.5) and alternations {a|b}
6. Generate: Get platform-optimized prompts

---

📝 Detail Levels Explained

Concise (~150-200 words)

• Essential details only
• Clear subject, basic setting, simple action
• Use when: You want compact prompts or API token savings

Moderate (~250-350 words)

• Good balance of detail
• Subject characteristics, environment description, basic aesthetics
• Use when: Standard detailed descriptions needed

Detailed (~400-500 words) - DEFAULT

• Rich, comprehensive description
• Full cinematography, specific lighting, camera work
• Use when: Professional video generation

Exhaustive (~600-1000 words)

• Maximum detail for cinematic quality
• Director-level descriptions, complete technical details
• Use when: Masterful, film-quality output required

---

🎨 Style Presets

Cinematic

Professional film quality with emphasis on lighting, composition, and smooth camera movements.

• Edge lighting, soft lighting, warm colors
• Professional framing, balanced composition

Surreal

Dreamlike, unusual scenes with artistic emphasis.

• Unexpected combinations, creative camera angles
• Artistic lighting, ethereal atmosphere

Action

High-energy motion and dynamic camera work.

• Fast movement, intense action
• Dynamic angles, motion blur

Stylized

Artistic interpretation with strong visual identity.

• Bold choices, distinctive look
• Creative freedom, unique aesthetics

Noir

Dark, moody film noir aesthetic.

• High contrast, dramatic shadows
• Low-key lighting, mystery atmosphere

Random

AI selects random aesthetic elements while respecting your core concept.

• Variety for batch generation
• Creative combinations

---

🔧 Configuration

Common Settings

LLM Configuration

• llm_backend: lm_studio or ollama
• model_name: Your loaded model (e.g., "llama3")
• api_endpoint: LLM server URL
• temperature: 0.1-2.0 (0.7 recommended, lower = focused, higher = creative)

Keywords

• positive_keywords: Comma-separated must-include terms (LoRA triggers, style terms)
• negative_keywords: Comma-separated terms to avoid

Output

• num_variations: Generate 1-3 variations
• save_to_file: Save prompts to disk with metadata
• filename_base: Base name for saved files

Node Outputs

All nodes return:

1. positive_prompt_1/2/3: Enhanced prompt variations
2. negative_prompt: Auto-generated platform-specific negatives
3. breakdown/settings_used: Detailed analysis of what was applied
4. status: Success messages, errors, file save location

---

🗂️ Maintaining Your Wildcard and LoRA Libraries

The Wildcard Composer and the LoRA Suggester are only as good as the lists and LoRA cards behind them. This is how to add to both, clean them up and keep their indexes current.

Wildcards

Where they live. Any wildcard folder ComfyUI knows (ComfyUI/models/wildcards, plus extra model paths such as A:\Models\wildcards). A few top-level folders are special:

Folder	What it is
_nsfw/	NSFW lists, used only at the matching NSFW level (_nsfw/suggestive/, _nsfw/nudity/ rate lower)
_prompts/	whole prompts, one per line: reached only through the Composer's Prompt line or __prompts/<pack>/...__
_generated/	lists the tools write
_dumped/	lists you dumped in the review; the engine skips it
mine/	your own lists: merges, lists made with the LLM

Your choices (stars, hidden lists, dropped lines, NSFW ratings, renames, button assignments) are kept in _catalog_user.yaml in the wildcard folder. The pack files themselves are never edited.

Adding lists. Copy the files in; nothing else is needed to use them. ComfyUI rescans the wildcard folders every 10 seconds, so a new list works in __name__ references and the Composer at once. The catalog behind Find, More lists and button ranking catches up by itself in the background each time ComfyUI starts: restart ComfyUI to refresh it.

Reviewing and cleaning: the wildcard review tool. Run tools\wildcard_review\wildcard_review.bat and open http://127.0.0.1:8766. (Close an old review window first: a server still running holds the port, and you would see the old version.) Per list you can:

• Keep / Dump / Same as / Merge (keys K / D / S / M; A takes the AI judge's call). Dump hides a list and never deletes it.
• Rename a list you keep, so it is easy to find in pills and pages. Its old name keeps working, so saved workflows are safe.
• Put it on Composer buttons, any number of them. + new button makes one of your own, under the heading you pick (Setting › Place, Subject › Clothing ...).
• Merge it with similar lists. A preview shows every combined line and which lists have it; untick any to leave them out. The result is written to mine/.
• Drop bad lines. Click a line and it is never picked again; this is saved at once. Drop matching drops every line with a word or regex. Suggest junk asks the local LLM which lines don't belong, and you confirm.
• Make new lists with the local LLM (LM Studio): + New list... from a description and a few samples, or Extend... to grow a short list. You check every line before it is saved to mine/.

Apply writes your decisions and does the renames and merges, showing a preview first. Move dumped files, in the same window, moves dumped files to _dumped/. Every rename, merge, move and new list writes a manifest; to undo one, run:

python tools\wildcard_review\server.py --undo tools\wildcard_review\manifests\<file>.json

The review tool reads the folder set in LIBRARY_ROOT at the top of tools/wildcard_review/server.py (default A:\Models\wildcards; the WCR_ROOT environment variable overrides it). The LLM features use LM Studio at http://localhost:1234; choose the model in the page's header menu.

Getting new lists into the review. The review shows lists the triage and the AI judge have seen. After adding packs, restart ComfyUI (so the catalog is current), then from this folder run the commands below. The judge needs LM Studio and only judges lists it has not seen.

..\..\..\python_embeded\python.exe -X utf8 dev\wildcard_triage.py
..\..\..\python_embeded\python.exe -X utf8 dev\wildcard_judge.py --all

Note: check how a fresh triage run treats lists you have since renamed, merged or moved to _dumped/ before re-running it on a reviewed library.

See docs/WILDCARD_GUIDE.md for the syntax and the folder rules, and docs/PROMPT_BUILDER_GUIDE.md for the Composer.

LoRAs (the LoRA Catalog)

The LoRA Suggester reads a catalog of LoRA "cards": what each LoRA does, its style and subject words. They are built by the companion LoRA Catalog tool, kept in its own folder (A:\LoRA_Catalog here). The path to its database is the lora_suggest section of config.json (default L:\Models\loras\_lora_catalog.sqlite).

After adding LoRAs, run refresh_new_loras.bat in the LoRA Catalog folder, with LM Studio running. Every step skips work already done, so it is safe to run any time. It writes refresh_new_loras.log.

Step	Command	What it does
1	run.py index	reads new or changed LoRA files (headers, sidecars, LoRA Manager data)
2	run.py crawl	fetches Civitai records for LoRAs that have none yet (internet; the run carries on if it fails)
3	run.py creators	looks up who made each LoRA (internet; the run carries on if it fails)
4	run.py cards --workers 3	the local LLM writes a card for every LoRA without one
5	run.py review-all	rebuilds the HTML review pages
6	run.py facets	rebuilds card_facets, the table the LoRA Suggester reads

refresh_new_loras.bat some/model-id writes the new cards with another LM Studio model. To redo cards that already exist, run step 4 by hand with --force (add --only Krea2/graphic-art to limit it to one folder), then step 6. Never run facets while a cards run is still writing.

Checking and correcting cards: the card editor. Run card_editor.bat in the LoRA Catalog folder and open http://127.0.0.1:8765. Your edits are kept apart from the generated cards and win over them. Reload the page to see newly written cards; after editing, run run.py facets so the Suggester sees your changes. The pick-list words live in vocab.yaml in the same folder.

📚 Documentation

Quick References

• QUICKSTART.md - Get started in 5 minutes
• QUICK_REFERENCE.md - Fast reference guide
• QUICK_REF_V17.md - Text-to-Image quick reference

Node Guides

• TXT2IMG_GUIDE.md - Text-to-Image node complete guide
• IMG2IMG_GUIDE.md - Image-to-Image workflows
• PROMPT_BUILDER_GUIDE.md - Wildcard Prompt / Character Composer panel
• SURREAL_GUIDE.md - Surreal Premise node and the surreal engine
• WILDCARD_GUIDE.md - Wildcard syntax, folders and the review tool
• NODE_COMPARISON.md - Which node to use when

Setup Guides

• LM_STUDIO_SETUP.md - LM Studio configuration
• VISION_BACKEND_GUIDE.md - Qwen3-VL setup for image analysis
• CONFIGURATION.md - Advanced configuration options

Platform & Technical

• PLATFORM_NOTES.md - Platform profiles and structured formats
• WAN_GUIDE_REFERENCE.md - Wan 2.2 video prompt guide

Updates & Fixes

• CHANGELOG.md - Complete version history

---

🆕 What's New in v2.2.0

Wildcard Prompt Composer — build a prompt section by section, with buttons

Two new builder nodes write a prompt one section at a time, with clickable pills instead of typed wildcard syntax:

• Wildcard Prompt Composer holds every section in one node — Format, Subject, Action, Setting, Light & atmosphere, Composition, Camera and Style & finish (plus optional Text) — shown in the order the chosen platform wants (Krea 2, FLUX.2, ...). Each [Pill] is a wildcard; the words around them are yours. Lock a section to hold it while the seed re-rolls, drag sections to reorder, and save the result as a template.
• Wildcard Character Composer is the same node limited to the Subject section, for a character you reuse across several prompts or branches.

See docs/PROMPT_BUILDER_GUIDE.md for the full panel reference.

Wildcard review tool — curate your library in the browser

A local browser tool for cleaning up and organising your wildcard lists. Run tools\wildcard_review\wildcard_review.bat and open http://127.0.0.1:8766. Working one list at a time, you can:

• Keep or rename a list; the old name keeps working through aliases, so saved workflows and templates don't break.
• Put it on Composer buttons (any number), or make your own button under the heading you pick.
• Merge similar lists, with a preview of every combined line so you can untick ones to leave out.
• Drop bad lines — one click removes a line from future picks; "Drop matching" clears a pattern across lists.
• Extend a short list with the local LLM, checking every line before it's saved.

Apply writes your decisions (renames, merges, moves) after showing a preview, and every change writes a manifest you can undo. The tool reads the folder set in LIBRARY_ROOT at the top of tools/wildcard_review/server.py (default A:\Models\wildcards); LLM features use LM Studio at http://localhost:1234.

See the Wildcards → Reviewing and cleaning section below and docs/WILDCARD_GUIDE.md for full details.

---

🆕 What's New in v1.8.0

Advanced Prompt Expander Node - Complete Redesign

3 Major Usability Improvements:

1. Operation Modes - Now you can modify existing prompts!

   • expand_from_idea - Original expansion behavior
   • refine_existing - Polish and improve prompts
   • modify_style - Change aesthetic while keeping subject
   • add_details - Add descriptive richness

2. Clear Detail Levels - No more confusing tiers!

   • Renamed to: Concise, Moderate, Detailed, Exhaustive
   • Added tooltips explaining each option
   • Removed confusing "auto" mode

3. Image Input Support - Image-to-video actually works!

   • Optional reference_image input
   • Qwen3-VL vision analysis integration
   • Automatic mode detection (text vs image-to-video)
   • Status shows when image is being used

Bonus: All video nodes now support emphasis (keyword:1.5) and alternation {a|b|c} syntax!

---

💡 Tips & Best Practices

For Best Results

1. Start Simple: Begin with basic prompts, let the LLM expand
2. Use Presets: They provide consistent, proven aesthetic directions
3. Experiment with Temperature: 0.7 is balanced, 0.3-0.5 for consistency, 0.8-1.2 for variety
4. Leverage Keywords: Add LoRA triggers and style terms to positive_keywords
5. Save Your Prompts: Enable save_to_file to build a library
6. Use Variations: Generate 3 variations to pick the best
7. Platform Matters: Choose the right platform for your model

Common Workflows

Quick Video Prompt

Node: Video Prompt Expander
Input: "cyberpunk street scene"
Preset: stylized
Detail: moderate

Professional Image Generation

Node: Text-to-Image Enhancer
Platform: Flux
Input: "portrait of a (warrior:1.3) with {red|blue|purple} armor"
Length: long
Genre: fantasy

Refine Existing Prompt

Node: Video Prompt Expander (Advanced)
Operation: refine_existing
Input: [Your existing 300-word prompt]
Detail: detailed

Image-to-Video with Direction

Node: Video Prompt Expander (Advanced)
Reference Image: [Your image]
Input: "Camera slowly dollies in while subject looks up"
Operation: expand_from_idea

---

🐛 Troubleshooting

LLM Connection Failed

• Check LLM backend is running (LM Studio or Ollama)
• Verify endpoint URL matches your LLM server
• Test in browser: http://localhost:1234/v1 (LM Studio) or http://localhost:11434 (Ollama)

Empty Output

• Check temperature isn't too high (>1.5)
• Verify model is loaded in LLM backend
• Check ComfyUI console for error messages

Emphasis Syntax Not Working

• Make sure you're using parentheses with colon: (keyword:1.5)
• Syntax is preserved in v1.6.1+ for video nodes, v1.7+ for image nodes

Image Analysis Not Working

• Install vision dependencies: pip install transformers accelerate huggingface_hub bitsandbytes
• Check Qwen3-VL model is downloaded
• See VISION_BACKEND_GUIDE.md

---

🤝 Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Test thoroughly
5. Submit a pull request

---

📄 License

Dual License:

• Non-Commercial: Free for personal/research use
• Commercial: Contact for licensing

See LICENSE for details.

---

🙏 Acknowledgments

• ComfyUI community for the amazing platform
• LM Studio and Ollama teams for excellent LLM backends
• Qwen3-VL team for the vision model
• All users who provided feedback and bug reports

---

📞 Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions

---

Version: 2.1.0
Last Updated: September 21, 2026
Author: Eric Rollei
Repository: https://github.com/EricRollei/Local_LLM_Prompt_Enhancer