README.md

---
base_model: Boogu/Boogu-Image-0.1-Turbo
base_model_relation: quantized
tags:
- comfyui
- gguf
- text-to-image
- boogu
---

# Boogu-Image-0.1-Turbo GGUF

<img src="https://cdn-uploads.huggingface.co/production/uploads/68761990332d15464ccc8dee/4BSudCXfjSA0ILq8HFcOE.png" width="300" height="300" />   <img src="https://cdn-uploads.huggingface.co/production/uploads/68761990332d15464ccc8dee/Vfo9RtMYbOK_k0fQuAcfT.png" width="300" height="300" />


### - V2 GGUFs are LIVE! 

### Required Text Encoder & VAE 

* **Text Encoder (Qwen3-VL):** You must use an FP8 scaled version of the Qwen3-VL encoder.
  * Download the FP8 encoder from the [Comfy-Org Boogu Repo](https://huggingface.co/Comfy-Org/Boogu-Image/tree/main/text_encoders).
* **VAE (Flux):** The Boogu pipeline utilizes the standard Flux VAE.
  * Download `flux1_vae_bf16.safetensors` from the [Comfy-Org Boogu Repo](https://huggingface.co/Comfy-Org/Boogu-Image/tree/main/vae)

### ⚠️ Why are there no Q2 or Q3 K-Quants?
Currently, this repo only provides **Flat Quants** (`Q4_0`, `Q4_1`, `Q5_0`, `Q5_1`, `Q8_0`). 

Standard K-quants (like `Q2_K` or `Q3_K_M`) require a hardcoded architectural mapping blueprint inside the `llama.cpp` source code. Because the Boogu/OmniGen architecture is brand new, those K-quant blueprints do not exist in the compiler yet. Flat quants bypass this requirement by forcing all 2D tensors to the target bit-depth. 

If you are running an 8GB VRAM setup (like an RTX 3070), the **Q4_0** is the recommended sweet spot for VRAM savings and quality.

### How to use in ComfyUI

**Prerequisite: Core Update Override**
Native support for the Boogu/OmniGen architecture has been merged from Pull Request #14523. If your Load Clip node doesnt have the Boogu architecture, you must manually fetch the PR into your ComfyUI installation.
1. Open a command prompt directly inside your `ComfyUI` folder.
2. Run the following commands to fetch and switch to the PR branch:
```cmd
   git fetch origin pull/14523/head:boogu-pr
   git checkout boogu-pr