Four ultra-lightweight, local-only nodes for ComfyUI:
🚀 Rebels Prompt Enhancer — curated Qwen3.5-4B text rewriter with a layered style system
🧪 Rebels Prompt Enhancer (Custom GGUF - experimental) — bring your own model, full control over every parameter
👁️ Rebels Image to Prompt (EXPERIMENTAL, NEED WORK) — vision-capable node that turns a reference image into a prompt
🔒 Rebels Prompt Locker — workflow gate that displays the passed-through prompt on its node face
No API keys, no external calls, 100% private. Aggressive VRAM cleanup so your diffusion model gets the GPU back immediately after the prompt is generated.
Features
Universal across all nodes
Local-Only. Everything runs on your machine via llama-cpp-python. Nothing leaves the box.
Aggressive VRAM Cleanup. Models fully unload after each call.
Prompt Lock + Cache. Toggle LIVE to iterate freely, then LOCKED to freeze the result you like. Locked = no model load on subsequent runs.
Full Diagnostic Output.thought_process output shows the assembled system prompt, raw model output, what was stripped, and the final result.
Layered Style System (text enhancers)
Three independent dropdowns compose into the system prompt at runtime:
Purpose — Image / Video / Edit. Controls subject framing.
Model Format — How the prompt should be structured for the model you're feeding.
Aesthetic — The visual style/vibe to apply.
This means hundreds of useful combinations from short dropdowns, all powered by the same 4B model.
Model Format Support
Format
Use For
Flux / Chroma
Natural-language Flux and Chroma models
Z-Image / Lumina-2
LLM-text-encoder models
HiDream
HiDream multi-encoder pipeline
SDXL
Tag + weight syntax for SDXL
SD 1.5
Tag + weight syntax for SD 1.5
Pony / Illustrious
Booru tags with score and rating tags
LTX Video
Motion-focused prose for LTX
Hunyuan / Wan Video
Cinematic motion prose
Universal Natural Language
Generic, works with most models
Aesthetic Library (22 styles)
Photorealistic · Cinematic Film · Anime/Manga · Studio Ghibli · Pixar/3D Animation · Comic Book · Concept Art · Oil Painting · Watercolor · Pencil Sketch · Cyberpunk · Steampunk · Fantasy · Sci-Fi · Horror/Dark · Vintage/Retro Film · Film Noir · Glamour/Editorial · Minimalist · Surreal/Dreamy · 3D Render/CGI · None (skip aesthetic injection)
Vision Support
The Image to Prompt node accepts any vision-capable GGUF + its paired mmproj projector. Supports multiple architectures via llama-cpp-python chat handlers:
LLaVA 1.5 / 1.6
Moondream
MiniCPM-V 2.6
NanoLLaVA
Qwen2.5-VL
Auto-detect picks the right handler from the filename, or override manually.
The Nodes
🚀 Rebels Prompt Enhancer (curated)
Uses Qwen3.5-4B locally with /no_think baked in to suppress reasoning mode. Output arrives clean.
Input
Purpose
raw_prompt
Your input idea
purpose
Image / Video / Edit
model_format
Output structure for your target model
aesthetic
Visual style to apply
precision
Efficiency (UD-IQ2, ~2GB) or Quality (UD-Q8, ~6.5GB)
seed
Randomize freely; ignored when locked
lock_in
LIVE or LOCKED
Outputs enhanced_prompt (clean string for the sampler) and thought_process (full diagnostic).
🧪 Rebels Prompt Enhancer (Custom GGUF)
(WARNING: this node is experimental and will sometimes provide bad results depending on gguf you use and format along with settings in the node. i suggest tweaking temperature and various settings until correct output is achieved.)
Same layered style system as the curated enhancer, but scans the folder for any .gguf and lets you pick. Adds full sampling controls.
Additional inputs vs the curated node:
Input
Purpose
model_file
Pick any GGUF found in the node folder
extra_instructions
Appended to the layered system prompt
system_prompt_override
If non-empty, replaces the layered prompt entirely (full manual mode)
append_no_think
Toggle the Qwen3 directive (only useful for Qwen3 family models)
n_gpu_layers
-1 = all on GPU, 0 = CPU only, N = partial offload
Look for abliterated versions from mlabonne or huihui-ai on HuggingFace
Larger models for users with more VRAM: use n_gpu_layers to control GPU offload at any model size
Vision models (for 👁️ Image to Prompt)
Vision setups need two files: the main GGUF and a paired mmproj-*.gguf (vision projector). Both go in the node folder.
Recommended starting picks:
Qwen3.5-4B-Uncensored (HauhauCS link above) — includes mmproj, works as both text AND vision model. Most efficient option if you're already using it for text.
Moondream2 — ~1.6B, tiny and fast, solid general captioning
Qwen2.5-VL-3B — modern vision model, fits 8GB VRAM cleanly
MiniCPM-V 2.6 — 8B, SOTA for small VLMs, best descriptive quality
Hardware Notes
The enhancers aggressively unload the model after each run, so VRAM only needs to fit the LLM during the rewrite — not alongside your diffusion model.
Approximate VRAM usage (model + 4k context)
Model
VRAM
Qwen3.5-4B UD-IQ2
~2 GB
Qwen3.5-4B UD-Q8
~6.5 GB
Qwen3.5-4B-Uncensored Q4_K_M
~3 GB
Moondream2
~1.5 GB
Qwen2.5-VL-3B Q4
~2.5 GB
MiniCPM-V 2.6 Q4
~5–6 GB
For bigger models
The Custom Enhancer's n_gpu_layers setting controls GPU offload:
-1 (default) — all layers on GPU (requires the model fits VRAM)
0 — pure CPU (slow but works for any model size, no VRAM needed)
N — partial offload, N layers on GPU and the rest on system RAM
This means users with bigger cards can drop in 30B+ models and run them natively, and anyone can experiment with partial offload for models that exceed their VRAM.
Usage
Find all four nodes under "Rebel AI" in the node menu.
[Load Image] → 👁️ Image to Prompt (Caption) → 🚀 Enhancer (apply Format + Aesthetic) → 🔒 Locker → KSampler
This chain captures the reference faithfully with the vision model, then runs that description through Qwen3.5-4B for the final styling — which often beats letting the vision model do both jobs at once, because text models are better wordsmiths than vision models.
The Lock workflow
Start the enhancer in LIVE with seed on randomize.
Queue runs and iterate — each run produces a different prompt because the seed changes.
When you find a prompt you love, toggle the enhancer to LOCKED. From then on it skips model loading and returns the cached prompt every run. Seed changes are ignored.
Independently, the Locker gates downstream execution — flip it to LOCKED IN to let the prompt flow to the sampler.
The two locks are intentionally separate:
Enhancer lock freezes the prompt (cache).
Locker halts the workflow (gate).
The cache lives in memory for the ComfyUI session and clears on restart. Changing any non-seed input while locked invalidates the cache and triggers a fresh generation under the new key.
Folder Structure
RebelsPromptEnhancer/
├── __init__.py
├── requirements.txt
├── README.md
├── *.gguf ← drop text model files here
├── mmproj-*.gguf ← vision projector files (paired with vision models)
└── web/
└── js/
└── rebels_locker_display.js
The web/js/ extension is what makes the Locker display its text on the node face. Without it, the Locker still functions but won't render the text visually.
