π¬ Overview
This custom ComfyUI node transforms prompt generation into a full cinematic directing system powered by Gemma 4.
Instead of writing basic prompts, this node acts as an AI Director, automatically generating:
Structured cinematic prompts (LTX 2.3 optimized)
Timeline-based scenes (perfect for video generation)
Camera direction, environment, subject behavior
Sound design layers
Genre-aware trailer logic (Netflix-style)
Whether you're creating short cinematic clips or full 20-second trailers, this node handles everything β from concept to final prompt.
π Key Features
π₯ Structured Cinematic Prompts
Clean sections:
[Camera][Environment][Subject][Timeline][Sound]
Optimized for LTX 2.3 video generation
π¬ Auto Director System
Fully automatic scene creation with multiple modes:
Smart Auto β balanced, high-quality results
Full Auto β complete scene generation
Chaos Auto β maximum creativity & variation
Generates automatically:
subjects & relationships
environments
camera style
actions & motion
sound design
π Subject Control
Choose or randomize:
Female subject
Male subject
Hetero couple
Lesbian couple
Or let AI decide
π Expanded Environment Database
Massive internal preset library:
Cinematic interiors
Urban & nightlife scenes
Nature & fantasy environments
Horror & liminal spaces
Action & disaster setups
Auto-selected or manually controlled.
π¬ Netflix-Style Trailer Mode (NEW)
Generate full 20-second cinematic trailers with:
5 structured timeline beats
Genre-based pacing
Dynamic camera progression
Sound design adapted to genre
Final hook shot
Supported Genres:
Psychological Horror
Sci-Fi Thriller
Action Spectacle
Dark Fantasy
Romantic Drama
Crime / Mystery
Dystopian / Survival
And many more
βοΈ Style Presets
Structured modes include:
Structured Romance
Structured Power
Structured Action
Structured Horror
Structured Dreamlike
Structured Realism
Structured POV
Structured Epic
Structured Minimal
π§ Why This Node?
This is not just a prompt generator.
It is a:
π Cinematic Prompt Engine
π AI Scene Director
π Video Pre-Production Tool
Designed specifically for:
LTX 2.3 workflows
Text-to-video pipelines
High-end cinematic AI generation
π Installation Guide (Step-by-Step)
1. Download the Node
Download the file:
gemma4_prompt_gen.py2. Place the File
Move it into:
ComfyUI/custom_nodes/Gemma4Prompt/If the folder doesnβt exist, create it manually.
3. Verify Structure
Your folder should look like:
ComfyUI/
βββ custom_nodes/
βββ Gemma4Prompt/
βββ __init__.py
βββ gemma4_prompt_gen.pyβ οΈ Do NOT rename __init__.py
β οΈ File name must be exactly: gemma4_prompt_gen.py
4. Install Dependencies
Make sure you have:
Python 3.10+
CUDA working (for GPU inference)
llama.cpp build with CUDA support
5. Setup llama-server
Download a CUDA-enabled build and run:
llama-server.exe -m YOUR_MODEL.gguf -ngl 60 --ctx-size 4096Make sure it's running on:
http://127.0.0.1:80806. Restart ComfyUI
Launch ComfyUI again and check:
Node appears as Gemma4PromptGen
No errors in console
7. Load Workflow
Use your existing workflow or connect:
Instruction input
Optional environment / subject
Output prompt β LTX pipeline
β‘ Recommended Settings
π¬ Cinematic Clip
Mode:
AUTOStyle:
StructuredAuto Director:
Smart Auto
π₯ Netflix Trailer
Style:
TrailerGenre:
Psychological Horror(example)Intensity:
BalancedorHigh ImpactSound:
Auto by Genre
π₯ Full Automatic Generation
Auto Director:
Full AutoSubject:
OffEnvironment:
LLM decides
π‘ Final Notes
Higher token limits = longer, richer prompts
Structured mode is best for LTX
Trailer mode is ideal for storytelling
π― Result
With this node, you move from:
β basic prompts
β‘οΈ to
π₯ AI-directed cinematic storytelling
If you like it, drop a like β
and share your generations β Iβd love to see what you create!
Description
FAQ
Comments (9)
Oh! My favorite! When someone uploads only two examples, and they're both bad and don't showcase what they're trying to offer. Only beaten by zero examples trustmebro posts.
Description doesn't state if this requires an API or if it's done locally...
only local
Description clearly says LLMC is running locally.
Is this an extended version of this Node here? Then at least mention it (for kindness), please.
similar but not the same, I took inspiration and improved it
We need start llama server manually or workflow doing this automatic ?
And where i choose lenght of video ? lol cant see
Great work thank you!
Do you think about making a version with 2 references images for FLF2V?
I often get an OOM during upscaling. I have 12 GB VRAM.