Qwen-Image 2.1 Turbo Edit (4–6 Step)

Requirements

• NVIDIA RTX 30XX / 40XX / 50XX GPU
• CUDA-compatible GPU — ~16 GB for int8, or 6-8 GB with a GGUF quant + the w4a8 encoder
• At least 20 GB free storage

Model downloads

• Viggle 6-step — Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r256.safetensors — loaded by default in this workflow, run at ~6 steps. By Viggle https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo/tree/main ⧉
• Alibaba-PAI 4-step — Qwen-Image-2.1-Fun-Acc-4Step.safetensors — the faster alternate, run at 4 steps. By Alibaba-PAI https://huggingface.co/alibaba-pai/Qwen-Image-2.1-Fun-Acc-LoRAs/tree/main/models ⧉
• qwen_image_2.1_int8_convrot.safetensors – ~7 GB, the default https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/diffusion_models ⧉
• qwen_image_2.1_Q5_K_M.gguf – low VRAM, ~4-7 GB by quant https://huggingface.co/Abiray/Qwen-Image-2.1-GGUF/tree/main ⧉
• qwen3vl_8b_int8_convrot.safetensors – alternate encoder https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/text_encoders ⧉
• qwen_image_2.1_vae_bf16.safetensors – required https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/vae ⧉
• Qwen/Qwen-Image-2.1 https://huggingface.co/Qwen/Qwen-Image-2.1 ⧉
• Alibaba-PAI https://huggingface.co/alibaba-pai/Qwen-Image-2.1-Fun-Acc-LoRAs/tree/main ⧉

Custom nodes

• ComfyUI Manager https://github.com/ltdrdata/ComfyUI-Manager ⧉
• ComfyUI-GGUF (for the GGUF loader path) https://github.com/city96/ComfyUI-GGUF ⧉
• rgthree-comfy (Image Comparer, bypasser) https://github.com/rgthree/rgthree-comfy ⧉

Setup steps

1. Download the free workflow .json with the button above.
2. Download the 8 model file(s) listed above into their ComfyUI model folders.
3. Install the 3 custom node(s) listed above via ComfyUI Manager, then fully restart ComfyUI.
4. Drag the workflow .json onto the ComfyUI canvas and confirm each loader node has the right model selected.
5. Run it. If a node shows red, update ComfyUI and restart before troubleshooting anything else.

Notes & tips

• ⚠️ Read This First - Turbo Caution: Image editing: the turbo LoRAs are consistently great here. This is what the workflow is built for, and it's reliable.; Text-to-image: it's hit-or-miss. Some prompts come out clean, others come out incomplete or smeared. I saw the same behavior on both LoRAs, so it isn't one bad file — the cause is unknown, and hopefully the creators tighten it up in a future version.
• The Two Turbo LoRAs: Viggle 6-step — Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r256.safetensors — loaded by default in this workflow, run at ~6 steps. By Viggle — Hugging Face Repo; Alibaba-PAI 4-step — Qwen-Image-2.1-Fun-Acc-4Step.safetensors — the faster alternate, run at 4 steps. By Alibaba-PAI — Hugging Face Repo
• Recommended Settings: Steps: 6 for the Viggle LoRA, 4 for the Alibaba-PAI LoRA; CFG: 1; Sampler: er_sde (euler also works); Scheduler: simple; LoRA strength: 1.0
• Load your source image through the Load Image node, and pick your model path (FP8/int8 or GGUF) with the Fast Groups Bypasser.
• Describe the edit in plain language — Qwen-Image 2.1 is instruction-driven, so "make the sky sunset orange" works better than keyword soup.
• For multi-image edits, you can supply up to 10 reference images; the output follows the first image's canvas.
• Keep custom resolution close to the resized source, or the edit can shift.
• Turbo is for edits — for reliable text-to-image, switch back to the full-step non-turbo model (see the caution above).

More resources

• Qwen-Image 2.1 All-in-One One-Click Installer (Patreon) https://www.patreon.com/TheLocalLab/posts/qwen-image-2-1-170300735 ⧉