Boogu-Image-Turbo_GGUFs

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.
• VAE (Flux): The Boogu pipeline utilizes the standard Flux VAE.
  • Download flux1_vae_bf16.safetensors from the Comfy-Org Boogu Repo

⚠️ 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:

   git fetch origin pull/14523/head:boogu-pr
   git checkout boogu-pr