Qwen-Image 2.1 GGUFs

GGUF:
https://huggingface.co/realrebelai/Qwen-Image-2.1_GGUFs/tree/main

GGUF quantizations of Qwen-Image 2.1 for use in ComfyUI with ComfyUI-GGUF.

These files are converted from the Comfy-Org Qwen-Image 2.1 BF16 diffusion model and preserve the native Comfy tensor layout. No additional tensor-name remapping is required after conversion.

This repository contains the diffusion model only. You still need the normal Qwen-Image 2.1 supporting models required by your ComfyUI workflow, such as the text encoder and VAE.

Quantizations

Planned / available quant ladder:

QuantRelative SizeNotesQ8_0Largest quantHighest precision quantized optionQ6_KLargeVery high quality / good precisionQ5_K_MMedium-LargeHigh qualityQ4_K_MMediumRecommended balance of size and qualityQ3_K_MSmallLower VRAM / more aggressive quantizationQ2_KSmallestMost aggressive quantization; quality may degrade

Quantization quality can vary by prompt, resolution, sampler, and workflow. Test lower-bit versions before relying on them for production work.

Source

BF16 diffusion model source:

Comfy-Org/Qwen-Image-2.1

diffusion_models/qwen_image_2.1_bf16.safetensors

Official / upstream Qwen-Image 2.1 model components and configuration are from the Qwen project.

Architecture

The GGUF files use:

general.architecture = qwen_image

The source checkpoint contains:

• 32 transformer blocks

• 4096 hidden dimension

• 32 attention heads

• 128 head dimension

• 64 input channels

• 64 output channels

The Comfy-Org BF16 checkpoint already uses the tensor naming/layout expected by ComfyUI, so the conversion preserves those names instead of remapping them to a separate llama.cpp-style naming scheme.

Precision handling

The repeating transformer blocks are quantized according to the selected GGUF format.

Small or sensitive non-repeating tensors are preserved at higher precision where appropriate, including components such as:

img_in
txt_in
time_text_embed
modulation
norm_out
proj_out

The conversion also stores comfy.gguf.orig_shape.* metadata where a tensor must be physically reshaped for GGUF quantization compatibility. ComfyUI-GGUF restores the logical tensor shape when loading.

ComfyUI Installation

Install or update ComfyUI-GGUF:

cd /d D:\AI_Tools\ComfyUI_windows_portable\ComfyUI\custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF

Then install its requirements with ComfyUI's embedded Python:

D:\AI_Tools\ComfyUI_windows_portable\python_embeded\python.exe -m pip install -r D:\AI_Tools\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-GGUF\requirements.txt

If you already have ComfyUI-GGUF installed, update it instead of cloning a second copy.

Model Placement

Place the GGUF diffusion model in your normal ComfyUI diffusion-model directory, for example:

ComfyUI/models/diffusion_models/

Then load it with:

Unet Loader (GGUF)

or the equivalent GGUF diffusion-model loader supplied by your installed ComfyUI-GGUF version.

Keep using the standard Qwen-Image 2.1 text encoder, VAE, conditioning, sampler, and workflow components around it.

Important

These are unofficial community quantizations.

GGUF support for image diffusion models is still evolving. Make sure ComfyUI and ComfyUI-GGUF are current enough to support qwen_image.

Lower-bit quantization can affect:

• fine detail

• typography

• prompt adherence

• edit fidelity

• anatomy

• texture consistency

If a lower quant produces obvious degradation, move up one quant level before changing the rest of your workflow.

Credits

• Qwen — Qwen-Image / Qwen-Image 2.1

• Comfy-Org — ComfyUI-compatible Qwen-Image 2.1 model packaging

• City96 / ComfyUI-GGUF — GGUF loading and quantization support for ComfyUI

• llama.cpp — GGUF quantization infrastructure

Quantizations / conversion work by RealRebelAI.

Links

• Qwen-Image 2.1 model files: https://huggingface.co/Comfy-Org/Qwen-Image-2.1

• ComfyUI-GGUF: https://github.com/city96/ComfyUI-GGUF