|
|
--- |
|
|
base_model: Casual-Autopsy/L3-Umbral-Mind-RP-v3.0-8B |
|
|
library_name: unsloth |
|
|
language: en |
|
|
tags: |
|
|
- llama-3 |
|
|
- stable-diffusion |
|
|
- flux |
|
|
- z-image-turbo |
|
|
- prompt-engineering |
|
|
- uncensored |
|
|
- gguf |
|
|
license: llama3 |
|
|
--- |
|
|
|
|
|
# ⚡ Bojro PromptMaster Uncensored v2 8B |
|
|
|
|
|
**Bojro PromptMaster Uncensored v2** is the official 8B parameter companion model for the **[Bojro Resolver App](https://github.com/bojrodev/Resolver-WebUI-Forge-Client)**. |
|
|
|
|
|
This model is a specialized fine-tune designed to serve as the high-intelligence prompt engine for the Bojro Resolver ecosystem. It runs on a PC backend (via LM Studio, Ollama, or llama.cpp) and bridges the gap between simple user intent and the complex technical requirements of **SDXL**, **Flux.1**, and **Z-Image Turbo**. |
|
|
|
|
|
## 📱 Official App (Client-Server) |
|
|
This model is built to be accessed via API by: |
|
|
👉 **[Bojro Resolver (Stable Diffusion Client for Android)](https://github.com/bojrodev/Resolver-Stable-Diffusion-Client-for-android)** |
|
|
|
|
|
--- |
|
|
|
|
|
## ⚠️ Critical Inference Settings |
|
|
To prevent tag looping (e.g., repeating "black nails" indefinitely), you **MUST** use these settings when hosting the model: |
|
|
|
|
|
| Parameter | Recommended Value | Why? | |
|
|
| :--- | :--- | :--- | |
|
|
| **Temperature** | `0.75` | Keeps creativity focused without hallucinating. | |
|
|
| **Repetition Penalty** | **`1.2`** | **CRITICAL:** Prevents tag-list loops in SDXL mode. | |
|
|
| **Top P** | `0.9` | Standard sampling for Llama 3. | |
|
|
|
|
|
--- |
|
|
|
|
|
## ⚙️ Required System Prompts |
|
|
|
|
|
To switch between the specialized engines, your client must send the specific System Prompt corresponding to the target generator. |
|
|
|
|
|
### 1️⃣ Target: SDXL / Pony |
|
|
*Use this for models that require structured Danbooru-style tags.* |
|
|
> **System Prompt:** |
|
|
> You are an expert SDXL Prompt Engineer. Your task is to expand the user's short concept into a high-quality, detailed generation prompt. Adapt the formatting, style, and tag usage to match the user's specified category (Anime, Pony, or Realistic). |
|
|
|
|
|
### 2️⃣ Target: Flux.1 |
|
|
*Use this for models utilizing the T5-XXL encoder.* |
|
|
> **System Prompt:** |
|
|
> You are the Flux Photographic Director. You translate ideas into immersive, natural language prose for the T5-XXL encoder. Describe scenes with spatial awareness, camera technicals, and lighting types. OUTPUT ONLY THE PROSE. |
|
|
|
|
|
### 3️⃣ Target: Z-Image Turbo (Qwen) |
|
|
*Use this for models utilizing the Qwen text encoder.* |
|
|
> **System Prompt:** |
|
|
> You are the Z-Image Narrative Engine. You specialize in dense, material-focused storytelling prompts for the Qwen text encoder. Focus on textures, atmospheric effects, and sensory details. OUTPUT ONLY THE NARRATIVE. |
|
|
|
|
|
--- |
|
|
|
|
|
## 💡 Usage Examples (API Response) |
|
|
|
|
|
### **SDXL Mode** |
|
|
**Output:** `score_9, score_8_up, score_7_up, 1cat, cyberpunk armor, glowing neon eyes, rain-slicked fur, futuristic city background, bokeh, 8k, unreal engine 5.` |
|
|
|
|
|
### **Flux Mode** |
|
|
**Output:** `A cinematic low-angle shot of a gritty cyberpunk cat prowling through a neon-lit alleyway. Rain drips from its metallic whiskers, reflecting the pink and blue holograms above. The depth of field is shallow.` |
|
|
|
|
|
### **Z-Image Turbo Mode** |
|
|
**Output:** `The sensation of cold, brushed steel against wet fur. A solitary feline figure clad in scavenged cyber-armor navigates the slick pavement. Textures of rusted metal contrast with the sleek, glowing blue circuitry.` |
|
|
|
|
|
--- |
|
|
|
|
|
## 📥 Deployment (PC Backend) |
|
|
|
|
|
This model is provided as a **GGUF** file optimized for PC hosting: |
|
|
|
|
|
1. **Host:** Load `BojroPromptMaster_uncensored_v2-8B-Q4_K_M.gguf` **(<- RECOMMENDED)** into LM Studio or any OpenAI-compatible API server. |
|
|
2. **Connect:** Point your **Bojro Resolver** Android app to your PC's local IP address (e.g., `http://192.168.1.XX:1234/v1`). |
|
|
3. **Configure:** Ensure the Repetition Penalty is set to **1.2** in your host settings. |
|
|
|
|
|
## 📜 License & Credits |
|
|
|
|
|
- **Base Model:** [Casual-Autopsy/L3-Umbral-Mind-RP-v3.0-8B](https://huggingface.co/Casual-Autopsy/L3-Umbral-Mind-RP-v3.0-8B) |
|
|
- **Finetuned using:** [Unsloth](https://github.com/unslothai/unsloth) |
|
|
- **License:** **Meta Llama 3 Community License**. |
