Last updated:
Qwen Image 2.1 Turbo in ComfyUI and on your own GPU
ComfyUI repackaged the Turbo weights within hours of the October 9 release, and community GGUF quants followed. This page lists the exact files, the sampler settings that matter, and how to pick a quant for your graphics card.
Files to download
| File (Comfy-Org/Qwen-Image-2.1) | What it is |
|---|---|
| diffusion_models/qwen_image_2.1_turbo_bf16.safetensors | Full 8-step Turbo checkpoint, BF16 |
| diffusion_models/qwen_image_2.1_turbo_int8_convrot.safetensors | INT8 quantized Turbo, for less VRAM |
| loras/qwen_image_2.1_turbo_lora_avg_rank_178_bf16.safetensors | Turbo as a LoRA, applied to the base Qwen-Image 2.1 model |
| Text encoder and VAE | Shared with Qwen-Image 2.1; use the files from the same repo |
Source: Comfy-Org repackaged weights
Source: ComfyUI Wiki release note
ComfyUI setup, step by step
- Update ComfyUI to a build with Qwen-Image 2.1 support.
- Download the Turbo diffusion model into ComfyUI/models/diffusion_models/, plus the Qwen-Image 2.1 text encoder and VAE into their folders.
- Open the official Qwen-Image 2.1 text-to-image workflow.
- Point the model loader at the Turbo file, or keep the base model and add the Turbo LoRA.
- Set the sampler to 8 steps and CFG 1. Start at 1024 × 1024 to check it works, then move to the 2K presets.
Kijai notes the Turbo sigma schedule is close to ComfyUI’s linear_quadratic scheduler with no shift. Qwen says schedules other than the stored one are untested.
Source: ComfyUI Wiki release note
Source: Comfy-Org repackaged weights
Settings that matter
| Setting | Turbo value |
|---|---|
| Steps | 8 |
| CFG | 1 (one forward pass per step) |
| Sampler | Euler, flow matching |
| Sigmas | 1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568, then 0 |
| Resolution | 2048 × 2048 native; 1024 × 1024 for quick tests |
Source: Hugging Face model card
Source: ComfyUI Wiki release note
Source: AtomicChat GGUF card
GGUF quants and which to pick
AtomicChat published GGUF files for stable-diffusion.cpp and measured how far each drifts from the BF16 original (LPIPS at 1024 × 1024, lower is closer).
- Q8_0 is the closest quantized option; use it if it fits.
- AD-Q4_K is the recommended 4-bit pick. Q3 and Q2 visibly change layout, faces and lettering.
- You also need the Qwen-Image-2.1 VAE and a Qwen3-VL-8B-Instruct GGUF text encoder (Q8_0 recommended). Quantizing the encoder also changes outputs.
- stable-diffusion.cpp needs a build from October 6, 2026 or newer for custom sigmas. Use 8 steps, CFG 1.0, Euler and the Turbo sigmas.
| File | Size | LPIPS vs BF16 |
|---|---|---|
| BF16 | 14.23 GB | 0 |
| Q8_0 | 7.59 GB | 0.037 |
| AD-Q6_K | 6.71 GB | 0.044 |
| AD-Q5_K | 5.40 GB | 0.076 |
| AD-Q4_K | 4.20 GB | 0.147 |
| AD-Q3_K | 3.60 GB | 0.206 |
| AD-Q2_K | 2.55 GB | 0.359 |
Source: AtomicChat GGUF card
How much VRAM you need
Qwen has not published a VRAM minimum for Turbo. These are third-party estimates for the base model, which has the same size.
| Your hardware | What to run (Unsloth estimate) |
|---|---|
| 6 GB VRAM | FP8 with offloading, under 2x slower |
| 12–16 GB VRAM | GGUF Q4_K_M at 1024 × 1024 |
| 24 GB VRAM | INT8 / FP8 weights |
| CPU only, 12–16 GB RAM | GGUF Q4_K_M with a 4-bit text encoder (slow) |
| Apple Silicon, 16 GB+ | GGUF with a native backend |
Turbo needs the same memory as the base model per step; it saves time because it runs 8 steps instead of 40.
Source: Unsloth Qwen-Image-2.1 docs
Diffusers with less VRAM
In Python, CPU offloading keeps only the active component on the GPU. It is slower but lets a smaller card run the BF16 checkpoint.
pip install "torch>=2.4.0"
pip install git+https://github.com/huggingface/diffusers.git
pip install "transformers>=5.17.0" accelerate pillow
# Less VRAM: keep only the active part of the model on the GPU (slower)
pipe = QwenImage21Pipeline.from_pretrained("Qwen/Qwen-Image-2.1-Turbo", dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()Full Python example on the main guide
Source: Qwen-Image-2.1 on GitHub
Source: Hugging Face model card
FAQ
What are the best ComfyUI settings for Qwen Image 2.1 Turbo?
8 steps, CFG 1, Euler sampling, with the Turbo sigma schedule (close to linear_quadratic with no shift).
Turbo checkpoint or Turbo LoRA?
The checkpoint is simpler. The LoRA lets you keep one base model on disk and switch between 40-step and 8-step generation.
Which GGUF should I download?
Q8_0 (7.59 GB) if it fits, otherwise AD-Q4_K (4.20 GB). Avoid Q3 and Q2 unless nothing else fits.
Can I run it on 8 GB of VRAM?
Qwen publishes no minimum. Unsloth estimates FP8 with offloading runs on 6 GB (under 2x slower) and a 4-bit GGUF needs 12–16 GB at 1024 × 1024. Start small, batch size 1.
Why do my Turbo images look wrong?
Usually the schedule: changing steps without the Turbo sigmas, or running CFG above 1. Reset to 8 steps, CFG 1 and the stored schedule.
Independent guide, not affiliated with Alibaba or the Qwen team. Prices, licences and file names change; check the linked official pages before you build on them.