Qwen-Image 2.1 Turbo — GGUFs by Rebel AI
The official 8-step Qwen-Image 2.1 Turbo diffusion model, converted to GGUF for lower-memory local inference in ComfyUI.
These are community-made quantizations by Rebel AI, based on the BF16 Turbo diffusion checkpoint distributed by Comfy-Org. The underlying model and Turbo distillation are by the Qwen team. This repository does not contain a separately trained model, nor does it require merging a Turbo LoRA.
Important: Turbo is an 8-step model, but the sampling schedule matters. The original Qwen Turbo Diffusers pipeline automatically loads the checkpoint's trained 8-step sigma schedule. A generic 8-step Euler/simple workflow is not necessarily equivalent. Use a Qwen-Image 2.1 Turbo-compatible ComfyUI workflow with the correct schedule. Also keep CFG at 1.0 as a starting point; do not stack another Turbo/acceleration LoRA on these already-distilled weights.
Downloads and variants
Get the available .gguf files from the Files tab. Choose a file based on your available memory and desired quality. Lower-bit quantization generally reduces model storage and weight-memory requirements, but can affect image quality, and the full workflow also uses memory for the text encoder, VAE, and activations.
Common GGUF quantization tiers include:
| Type | Intended trade-off |
|---|---|
Q3_K_S | Smaller weight footprint; greater risk of quality loss |
Q4_K_S | Lower-memory balanced option |
Q5_K_S | More precision, larger download |
Q6_K | Higher precision, larger download |
Only variants actually present in the Files tab are released. Mixed-precision presets may retain higher-precision tensors. A filename alone does not establish the actual tensor-type breakdown; refer to the file's GGUF tensor metadata. Image quality, VRAM use, and performance have not been benchmarked for every variant.
Install in ComfyUI
- Update ComfyUI to a recent build with native Qwen-Image 2.1 support.
- Install a compatible GGUF loader with Qwen-Image 2.1 support, such as molbal/ComfyUI-GGUF. Restart ComfyUI after installation.
- Download one of this repository's
.ggufdiffusion models and place it inComfyUI/models/diffusion_models/. - Download the corresponding Qwen-Image 2.1 text encoder and VAE from Comfy-Org/Qwen-Image-2.1, if you don't already have them.
- Open a Qwen-Image 2.1 Turbo-compatible text-to-image or image-edit workflow. Use Unet Loader (GGUF) (the label may differ by loader version) in place of the regular diffusion-model loader, select your GGUF, and keep the remaining Qwen-Image 2.1 components connected.
Model locations
ComfyUI/
└── models/
├── diffusion_models/
│ └── [your Qwen-Image-2.1-Turbo .gguf file]
├── text_encoders/
│ └── qwen3vl_8b_int8_convrot.safetensors # or another compatible Qwen3-VL 8B encoder
└── vae/
└── qwen_image_2.1_vae_bf16.safetensors
- Text encoders: Comfy-Org text_encoders
- VAE: qwen_image_2.1_vae_bf16.safetensors
- General native workflow documentation: ComfyUI Qwen-Image 2.1 guide
This repository supplies diffusion-model GGUF weights only. The text encoder and VAE are separate downloads; the GGUF is not a complete standalone image-generation pipeline.
Turbo sampling guidance
| Setting | Guidance |
|---|---|
| Steps | 8, using the Turbo checkpoint's trained sigma schedule |
| CFG | 1.0 as the starting point |
| Sampler | Follow the matching Turbo workflow; do not assume a generic scheduler reproduces the official sigmas |
| Resolution | Qwen-Image 2.1 supports multiple aspect ratios and high-resolution output; memory use rises with resolution |
| Extra Turbo LoRA | Not needed — these are quantizations of the full Turbo checkpoint |
The official model's 8-step schedule is stored in its Diffusers pipeline configuration, not in the .gguf weight tensors. Loading only this GGUF does not automatically transfer that scheduler configuration into ComfyUI. If results look washed out, muddy, or unusually distorted, check the Turbo schedule and workflow before blaming the quantization.
Qwen-Image 2.1 supports text-to-image and image editing. Editing additionally requires the corresponding Qwen-Image 2.1 image-conditioning workflow and its reference-image inputs; simply connecting a GGUF loader is not enough.
About these conversions
- Base architecture: Qwen-Image 2.1, approximately 7B diffusion-model parameters.
- Starting weights:
diffusion_models/qwen_image_2.1_turbo_bf16.safetensorsfrom Comfy-Org's official ComfyUI repack. - Method: GGUF weight conversion and quantization; no model retraining or LoRA merge.
- Scope: Diffusion-model weights only; text encoder, VAE, and scheduler/workflow are separate.
- Validation: GGUF loading and image quality should be checked for each exact file. Performance and quality vary with the quantization type, backend, settings, and hardware.
A native ComfyUI W4A8 ConvRot .safetensors checkpoint is a different format, requiring the appropriate native ComfyUI loader rather than a GGUF loader. Do not treat W4A8 and GGUF as interchangeable.
Credits and license
- Original architecture and Turbo model: Qwen / Qwen-Image-2.1-Turbo
- ComfyUI BF16 source checkpoint: Comfy-Org / Qwen-Image-2.1
- GGUF tooling and runtime: city96/ComfyUI-GGUF and molbal/ComfyUI-GGUF
- GGUF conversions and distribution: Rebel AI
The underlying model is distributed under the Qwen Research License Agreement. The original model license and terms continue to apply to these quantized derivatives. Please review those terms before use or redistribution.
These are unofficial community quantizations, not an official Qwen team release.