realrebelai/Viggle_Qwen-Image-2.1-Turbo_GGUFs

Viggle Qwen-Image 2.1 Turbo GGUFs

Mixed-precision GGUF quantizations of Viggle/Qwen-Image-2.1-viggle-turbo for use with ComfyUI + ComfyUI-GGUF.

These files are built from Viggle's full fine-tuned transformer:

transformer/diffusion_pytorch_model.safetensors

They are not the separate Viggle LoRA. Do not stack the Viggle Turbo LoRA on top of these GGUFs.

This repository contains the diffusion transformer only. You still need the normal Qwen-Image 2.1 text encoder, VAE, and workflow components.

Important compatibility / fidelity note

The upstream Viggle transformer and the source safetensors used for conversion were verified byte-for-byte identical by SHA256:

a011a284b2dcb4b5f73aa7041b770d6bbff601bbb0850f3b6d49740b567b24e5

However, in local testing, the GGUF runtime did not reproduce the upstream diffusion_pytorch_model.safetensors output exactly even at BF16-equivalent GGUF precision.

The GGUFs are functional, but composition, identity, clothing/details, and fine structure can differ from the upstream transformer with the same prompt and seed.

This mismatch was observed before lower-bit quantization, so it should not be interpreted as Q4/Q3/etc. quantization damage alone.

These files are provided as practical low-VRAM community conversions while Qwen-Image 2.1 GGUF support continues to mature.

Recommended Viggle Turbo settings

Viggle Turbo is a distilled 4-step model.

For the ComfyUI workflow used to test these files:

Steps:      4
CFG:        1.0
Sampler:    euler
Scheduler:  simple
Negative:   blank
Denoise:    1.0

The upstream Viggle release uses a FlowMatch Euler scheduler configuration intended for the 4-step student.

Quant ladder

These builds use the same Qwen-Image 2.1 HQv3 mixed-precision policy used in my Qwen-Image 2.1 GGUF work.

File	Base quant	Attention Q/K/V/Out	img_mlp.out	img_mlp.gate_up
Q8_0-HQv3	Q8_0	Q8_0	Q8_0	Q8_0
Q6_K-HQv3	Q6_K	Q8_0	Q8_0	Q8_0
Q5_K_M-HQv3	Q5_K_M	Q8_0	Q8_0	Q6_K
Q4_K_M-HQv3	Q4_K_M	Q8_0	Q8_0	Q5_K
Q3_K_M-HQv3	Q3_K_M	Q6_K	Q6_K	Q4_K
Q2_K-HQv3	Q2_K	Q5_K	Q5_K	Q3_K

The following top-level Qwen modules are retained at higher precision where applicable:

img_in
txt_in
time_text_embed
modulation
norm_out
proj_out

Small/norm tensors also remain high precision where required by the GGUF conversion path.

Build policy

Every quantization rung is generated directly from the BF16 GGUF master.

The ladder is not cascaded:

BF16 -> Q8
BF16 -> Q6
BF16 -> Q5
BF16 -> Q4
BF16 -> Q3
BF16 -> Q2

It is never:

BF16 -> Q8 -> Q6 -> Q5 -> Q4 ...

This avoids compounding quantization error from one rung into the next.

Architecture

For broad compatibility with current ComfyUI-GGUF installations, published files use:

general.architecture = qwen_image

Qwen-Image 2.1 itself uses the newer Qwen Image 2.1 transformer implementation upstream. Current GGUF support is still evolving, which is why the fidelity note above is important.

Installation

Install ComfyUI-GGUF and place the GGUF file in your normal diffusion-model directory, commonly:

ComfyUI/models/diffusion_models/

or, depending on your setup:

ComfyUI/models/unet/

Load it with the GGUF diffusion / UNet loader.

Use your existing Qwen-Image 2.1 supporting models around it.

Supporting models

You still need the normal Qwen-Image 2.1 components, including:

• Qwen3-VL text encoder
• Qwen-Image 2.1 VAE
• Qwen-Image 2.1-compatible conditioning/workflow nodes

Which quant should I use?

For most low-VRAM users, start with:

Qwen-Image-2.1-viggle-turbo-Q4_K_M-HQv3.gguf

If you have more memory, try Q5_K_M, Q6_K, or Q8_0.

Q3_K_M and Q2_K are more aggressive and may show additional losses in:

• faces
• anatomy
• typography
• small texture
• prompt adherence
• edit fidelity

Source models

Viggle Turbo:

https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo

Base Qwen-Image 2.1:

https://huggingface.co/Qwen/Qwen-Image-2.1

Related RealRebelAI GGUFs

Base Qwen-Image 2.1 GGUFs:

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

This repository:

https://huggingface.co/realrebelai/Viggle_Qwen-Image-2.1-Turbo_GGUFs

Credits

• Viggle — Qwen-Image 2.1 Viggle Turbo distillation / fine-tuned transformer
• Qwen — Qwen-Image 2.1
• City96 / ComfyUI-GGUF — GGUF loading infrastructure
• llama.cpp / ggml — GGUF quantization infrastructure
• RealRebelAI — conversion, mixed-precision quantization, testing, and packaging

License

These are derivative quantizations of the upstream model and do not replace or alter the upstream license.

Review the licenses and terms of Viggle/Qwen-Image-2.1-viggle-turbo and Qwen/Qwen-Image-2.1 before redistribution or commercial use.