中文Sign up

realrebelai/Boogu-Image-Edit-Turbo_GGUFs

🤗 Hugging Face sourceapache-2.042 GBGGUF✓ 5 checksumsupdated today
Submit in one command

Run it next to your model folder. It makes the torrent, checks your files against Hugging Face, and submits it. You just start seeding and paste your key from your account. It only reads your files and never changes them. Read the script first if you like.

curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo realrebelai/Boogu-Image-Edit-Turbo_GGUFs ./model-folder
Needs a seeder →

Boogu-Image-0.1-Edit-Turbo — GGUF (Flat Quants)

GGUF quantizations of Boogu/Boogu-Image-0.1-Edit-Turbo, the distilled (Turbo) reference-image edit variant of the Boogu-Image family. Quantized for low-VRAM ComfyUI use — an 8GB card (RTX 3070 class) with 16GB system RAM can run these.

Quantized by realrebelai. These are the DiT only — you supply the text encoder and VAE separately (see below).

Files

File Quant Size
boogu-edit-turbo-dit-Q8_0.gguf Q8_0 ~11.6 GB
boogu-edit-turbo-dit-Q5_1.gguf Q5_1 ~8.64 GB
boogu-edit-turbo-dit-Q5_0.gguf Q5_0 ~8.04 GB
boogu-edit-turbo-dit-Q4_1.gguf Q4_1 ~7.44 GB
boogu-edit-turbo-dit-Q4_0.gguf Q4_0 ~6.84 GB

On 8GB VRAM, Q4_0 is the recommended sweet spot for the balance of VRAM savings and quality. Step up to Q5_1 or Q8_0 if you have the headroom and want maximum fidelity.

Why only flat quants (Q4_0 / Q4_1 / Q5_0 / Q5_1 / Q8_0)?

This repo provides flat quants only. Standard K-quants (Q2_K, Q3_K_M, etc.) require a hardcoded architectural mapping blueprint inside the llama.cpp source. 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, so they quantize cleanly where K-quants would fall back to near-full precision.

Required components (not included here)

Boogu will not run with standard SD or Flux encoders. You must download the specific text encoder and VAE:

  • Text Encoder (Qwen3-VL): the FP8 scaled Qwen3-VL encoder from the Comfy-Org Boogu repo. In your CLIPLoader, set type = boogu.
  • VAE (Flux): flux1_vae_bf16.safetensors from the Comfy-Org Boogu repo.

⚠️ Most "it looks low-res / soft / noisy" reports come from loading the wrong encoder or VAE (e.g. a different Qwen3-VL size), or from the CLIPLoader type not being set to boogu. Verify these two files before reporting an issue.

Prerequisite: Core Update (PR #14523)

Native support for the Boogu/OmniGen architecture was merged in Pull Request #14523. If your Load CLIP node has no boogu architecture option, fetch the PR into your ComfyUI install. Open a command prompt inside your ComfyUI folder:

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

Install

  1. Download one boogu-edit-turbo-dit-*.gguf and place it in ComfyUI/models/unet/.
  2. Download the Qwen3-VL FP8 encoder → ComfyUI/models/text_encoders/ (or clip/).
  3. Download flux1_vae_bf16.safetensors → ComfyUI/models/vae/.
  4. Load the DiT with the Unet Loader (GGUF), the encoder with CLIPLoader (type = boogu), and the VAE with Load VAE.
  5. As an edit model, feed your reference image into the workflow's edit/reference-image input.

Notes

  • Turbo is the distilled variant — run it at its reduced step count (follow the step/CFG guidance on the base Edit Turbo model card); the full 50-step schedules used for non-Turbo models are unnecessary here.
  • These files are the diffusion transformer only. The encoder and VAE are shared across the Boogu-Image family — if you already run Boogu Base or Turbo, you have them.

Quantized and published by realrebelai. Boogu-Image is created by Boogu; all credit for the base model to the original authors. Released under Apache-2.0, matching the base model license.