<div data-youtube-video><iframe width="640" height="480" allowfullscreen="true" autoplay="false" disablekbcontrols="false" enableiframeapi="false" endtime="0" ivloadpolicy="0" loop="false" modestbranding="false" origin playlist src="https://www.youtube.com/embed/Yi8pbWyr4zA?rel=1" start="0"></iframe></div><h1 id="rebels-prompt-enhancer-nodes-for-comfyui">Rebels Prompt Enhancer Nodes for ComfyUI</h1>WORK IN PROGRESS - More features coming soon!VERSION 2 Now LIVE!<h3 id="you-can-download-and-extract-the-zip-and-run-the-requirements-but-you-must-follow-the-dependency-installation-instructions-in-the-step-by-step-instructions-below.-there-are-build-dependencies-that-are-required-so-follow-the-installation-steps-even-if-you-just-download-the-zip-file.-you-would-just-ignore-the-git-clone-command-if-you-download-the-zip.">You can download and extract the ZIP and run the requirements but you MUST follow the dependency installation instructions in the step-by-step instructions below. There are BUILD DEPENDENCIES that are required so follow the installation steps even if you just download the ZIP file. you would just ignore the git clone command if you download the ZIP.</h3>Four ultra-lightweight, local-only nodes for ComfyUI:<ul><li>🚀 Rebels Prompt Enhancer — curated Qwen3.5-4B text rewriter with a layered style system</li><li>🧪 Rebels Prompt Enhancer (Custom GGUF - experimental) — bring your own model, full control over every parameter</li><li>👁️ Rebels Image to Prompt (EXPERIMENTAL, OUT OF ORDER) — vision-capable node that turns a reference image into a prompt</li><li>🔒 Rebels Prompt Locker — workflow gate that displays the passed-through prompt on its node face</li></ul>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.<hr /><h2 id="features">Features</h2><h3 id="universal-across-all-nodes">Universal across all nodes</h3><ul><li>Local-Only. Everything runs on your machine via <code>llama-cpp-python</code>. Nothing leaves the box.</li><li>Aggressive VRAM Cleanup. Models fully unload after each call.</li><li>Prompt Lock + Cache. Toggle LIVE to iterate freely, then LOCKED to freeze the result you like. Locked = no model load on subsequent runs.</li><li>Full Diagnostic Output. <code>thought_process</code> output shows the assembled system prompt, raw model output, what was stripped, and the final result.</li></ul><h3 id="layered-style-system-(text-enhancers)">Layered Style System (text enhancers)</h3>Three independent dropdowns compose into the system prompt at runtime:<ul><li>Purpose — Image / Video / Edit. Controls subject framing.</li><li>Model Format — How the prompt should be structured for the model you're feeding.</li><li>Aesthetic — The visual style/vibe to apply.</li></ul>This means hundreds of useful combinations from short dropdowns, all powered by the same 4B model.<h3 id="model-format-support">Model Format Support</h3>FormatUse ForFlux / ChromaNatural-language Flux and Chroma modelsZ-Image / Lumina-2LLM-text-encoder modelsHiDreamHiDream multi-encoder pipelineSDXLTag + weight syntax for SDXLSD 1.5Tag + weight syntax for SD 1.5Pony / IllustriousBooru tags with score and rating tagsLTX VideoMotion-focused prose for LTXHunyuan / Wan VideoCinematic motion proseUniversal Natural LanguageGeneric, works with most models<h3 id="aesthetic-library-(22-styles)">Aesthetic Library (22 styles)</h3>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)<h3 id="vision-support">Vision Support</h3>The Image to Prompt node accepts any vision-capable GGUF + its paired <code>mmproj</code> projector. Supports multiple architectures via <code>llama-cpp-python</code> chat handlers:<ul><li>LLaVA 1.5 / 1.6</li><li>Moondream</li><li>MiniCPM-V 2.6</li><li>NanoLLaVA</li><li>Qwen2.5-VL</li></ul>Auto-detect picks the right handler from the filename, or override manually.<hr /><h2 id="the-nodes">The Nodes</h2><h3 id="rebels-prompt-enhancer-(curated)">🚀 Rebels Prompt Enhancer (curated)</h3>Uses Qwen3.5-4B locally with <code>/no_think</code> baked in to suppress reasoning mode. Output arrives clean.InputPurpose<code>raw_prompt</code>Your input idea<code>purpose</code>Image / Video / Edit<code>model_format</code>Output structure for your target model<code>aesthetic</code>Visual style to apply<code>precision</code>Efficiency (UD-IQ2, ~2GB) or Quality (UD-Q8, ~6.5GB)<code>seed</code>Randomize freely; ignored when locked<code>lock_in</code>LIVE or LOCKEDOutputs <code>enhanced_prompt</code> (clean string for the sampler) and <code>thought_process</code> (full diagnostic).<h3 id="rebels-prompt-enhancer-(custom-gguf)">🧪 Rebels Prompt Enhancer (Custom GGUF)</h3>(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 <code>.gguf</code> and lets you pick. Adds full sampling controls.Additional inputs vs the curated node:InputPurpose<code>model_file</code>Pick any GGUF found in the node folder<code>extra_instructions</code>Appended to the layered system prompt<code>system_prompt_override</code>If non-empty, replaces the layered prompt entirely (full manual mode)<code>append_no_think</code>Toggle the Qwen3 directive (only useful for Qwen3 family models)<code>n_gpu_layers-1</code> = all on GPU, <code>0</code> = CPU only, <code>N</code> = partial offload<code>n_ctx</code>, <code>max_tokens</code>, <code>temperature</code>, <code>top_p</code>, <code>repeat_penalty</code>Sampling controls<h3 id="rebels-image-to-prompt-(experimental-out-of-order)">👁️ Rebels Image to Prompt (EXPERIMENTAL, OUT OF ORDER)</h3>Vision-capable node. Takes a reference image and outputs a prompt suitable for downstream generation.InputPurpose<code>image</code>ComfyUI IMAGE input<code>model_file</code>Vision-capable GGUF from the folder<code>mmproj_file</code>Paired mmproj projector file<code>chat_handler</code>Auto-detect, or pick (LLaVA 1.5/1.6, Moondream, MiniCPM-V 2.6, NanoLLaVA, Qwen2.5-VL)<code>vision_task</code>Caption / Caption + Format / SD Tags / Pose &amp; Anatomy / Custom<code>model_format</code>Applied when vision_task = "Caption + Format"<code>aesthetic</code>Applied when vision_task = "Caption + Format"<code>custom_instruction</code>Used when vision_task = "Custom Instruction"Standard sampling + GPU controls + lockSame as Custom GGUF enhancerOutputs <code>image_prompt</code> (clean) and <code>thought_process</code> (full diagnostic).<h3 id="rebels-prompt-locker">🔒 Rebels Prompt Locker</h3>Workflow gate that also displays the prompt on its node face.InputPurpose<code>text_input</code>String to gate (typically <code>enhanced_prompt</code> or <code>image_prompt</code>)<code>lock_in_prompt</code>PAUSED halts with ValueError; LOCKED IN passes throughThe on-node text display requires the bundled <code>web/js/rebels_locker_display.js</code> extension to be present.<hr /><h2 id="installation">Installation</h2><ol><li>Navigate to your ComfyUI <code>custom_nodes</code> folder:<pre><code>cd ComfyUI/custom_nodes
</code></pre></li><li>Clone this repository:<pre><code>git clone https://github.com/RealRebelAI/RebelsPromptEnhancer.git
</code></pre></li><li>Install dependencies.ComfyUI Portable (open cmd inside the cloned folder):<pre><code>../../python_embeded/python.exe -m pip install -r requirements.txt
</code></pre>ComfyUI Desktop:<pre><code>pip install -r requirements.txt
</code></pre>If you hit build errors, upgrade your build chain first:<pre><code>..\..\python_embeded\python.exe -m pip install --upgrade pip setuptools wheel scikit-build-core
</code></pre></li><li>Download model file(s) from the Models section below and drop them directly in this node pack's folder.</li><li>Restart ComfyUI.</li></ol><hr /><h2 id="models">Models</h2>All models are GGUF format and live directly in the node pack folder. Download only what you'll use.<h3 id="curated-text-model-(for-rebels-prompt-enhancer)">Curated text model (for 🚀 Rebels Prompt Enhancer)</h3>Repo: <a target="_blank" rel="ugc" href="https://huggingface.co/unsloth/Qwen3.5-4B-GGUF">https://huggingface.co/unsloth/Qwen3.5-4B-GGUF</a>PrecisionFileSizeEfficiency (UD-IQ2)<code>Qwen3.5-4B-UD-IQ2_XXS.gguf</code> (or any <code>UD-IQ2_*</code>)1.5–1.8 GBQuality (UD-Q8)<code>Qwen3.5-4B-UD-Q8_K_XL.gguf</code>5.95 GBThe node searches by filename substring (<code>qwen3.5-4b</code> + <code>ud-iq2</code> or <code>ud-q8</code>), so renames work as long as the key tokens are intact.<h3 id="custom-text-models-(for-custom-gguf-enhancer)">Custom text models (for 🧪 Custom GGUF Enhancer)</h3>Drop any GGUF text model in the folder and it'll show up in the dropdown. Examples:<ul><li>Standard models: Any architecture llama.cpp supports — Llama, Mistral, Phi, Gemma, Qwen, etc.</li><li>Uncensored fine-tunes (for NSFW prompt rewriting where standard models refuse or sanitize):<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive">https://huggingface.co/HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive</a> — Q4_K_M at 2.71 GB, also includes mmproj so it doubles as a vision model</li><li>Look for <code>abliterated</code> versions from <code>mlabonne</code> or <code>huihui-ai</code> on HuggingFace</li></ul></li><li>Larger models for users with more VRAM: use <code>n_gpu_layers</code> to control GPU offload at any model size</li></ul><h3 id="vision-models-(for-image-to-prompt)-(out-of-order)">Vision models (for 👁️ Image to Prompt) (OUT OF ORDER)</h3>Vision setups need two files: the main GGUF and a paired <code>mmproj-*.gguf</code> (vision projector). Both go in the node folder.Recommended starting picks:<ul><li>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.</li><li>Moondream2 — ~1.6B, tiny and fast, solid general captioning</li><li>Qwen2.5-VL-3B — modern vision model, fits 8GB VRAM cleanly</li><li>MiniCPM-V 2.6 — 8B, SOTA for small VLMs, best descriptive quality</li></ul><hr /><h2 id="hardware-notes">Hardware Notes</h2>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.<h3 id="approximate-vram-usage-(model-+-4k-context)">Approximate VRAM usage (model + 4k context)</h3>ModelVRAMQwen3.5-4B UD-IQ2~2 GBQwen3.5-4B UD-Q8~6.5 GBQwen3.5-4B-Uncensored Q4_K_M~3 GBMoondream2~1.5 GBQwen2.5-VL-3B Q4~2.5 GBMiniCPM-V 2.6 Q4~5–6 GB<h3 id="for-bigger-models">For bigger models</h3>The Custom Enhancer's <code>n_gpu_layers</code> setting controls GPU offload:<ul><li><code>-1</code> (default) — all layers on GPU (requires the model fits VRAM)</li><li><code>0</code> — pure CPU (slow but works for any model size, no VRAM needed)</li><li><code>N</code> — partial offload, N layers on GPU and the rest on system RAM</li></ul>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.<hr /><h2 id="usage">Usage</h2>Find all four nodes under "Rebel AI" in the node menu.<h3 id="basic-text-wiring">Basic text wiring</h3><pre><code>[Your Prompt Text] → raw_prompt → 🚀 Enhancer → enhanced_prompt → 🔒 Locker → KSampler
 → thought_process → Preview-As-Text (optional)
</code></pre><h3 id="image-to-prompt-wiring">Image-to-prompt wiring</h3><pre><code>[Load Image] → image → 👁️ Image to Prompt → image_prompt → 🔒 Locker → KSampler
 → thought_process → Preview-As-Text (optional)
</code></pre><h3 id="chained-(image-faithful-caption-styled-prompt)">Chained (image → faithful caption → styled prompt)</h3><pre><code>[Load Image] → 👁️ Image to Prompt (Caption) → 🚀 Enhancer (apply Format + Aesthetic) → 🔒 Locker → KSampler
</code></pre>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.<h3 id="the-lock-workflow">The Lock workflow</h3><ol><li>Start the enhancer in LIVE with seed on <code>randomize</code>.</li><li>Queue runs and iterate — each run produces a different prompt because the seed changes.</li><li>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.</li><li>Independently, the Locker gates downstream execution — flip it to <code>LOCKED IN</code> to let the prompt flow to the sampler.</li></ol>The two locks are intentionally separate:<ul><li>Enhancer lock freezes the prompt (cache).</li><li>Locker halts the workflow (gate).</li></ul>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.<hr /><h2 id="folder-structure">Folder Structure</h2><pre><code>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
</code></pre>The <code>web/js/</code> 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.<hr />Built by Rebel AI.

Rebels Prompt Enhancer

a simple prompt enhancer built with a qwen 3.5-4b gguf backbone for low vram prompt enhancement.

Files with SHA256: b9198f948859421fc85c583b9cc3a9d947af7c7a96c092d5a495cd4d99a7e520