--- title: Qwen-Image-2.1-viggle-4-steps-turbo-GGUF canonical_url: "https://www.modelscope.cn/models/Abiray/Qwen-Image-2.1-viggle-4-steps-turbo-GGUF" md_url: "https://www.modelscope.cn/models/Abiray/Qwen-Image-2.1-viggle-4-steps-turbo-GGUF.md" repository: Abiray/Qwen-Image-2.1-viggle-4-steps-turbo-GGUF last_updated: 2026-09-23 license: other pipeline_tag: text-to-image-synthesis tasks: - text-to-image-synthesis base_model: - Viggle/Qwen-Image-2.1-viggle-turbo base_model_relation: quantized library_name: - lora - gguf - pytorch frameworks: - pytorch downloads: 590 stars: 2 tags: - gguf - text-to-image - image-to-image - comfyui - distilled - turbo - 4-step - qwen - qwen-image --- # Qwen-Image-2.1-viggle-4-steps-turbo-GGUF > Qwen-Image-2.1-viggle-4-steps-turbo-GGUF - Abiray 在 ModelScope 开源的模型。Qwen-Image-2.1 Viggle Turbo GGUF (4-Step Distilled) Abiray/Qwen-Image-2.1-viggle-4-steps-turbo-GGUF 是 ModelScope 魔搭社区上的text-to-image-synthesis模型，采用 other 许可，基于 Viggle/Qwen-Image-2.1-viggle-turbo 构建。 - **Repository**: Abiray/Qwen-Image-2.1-viggle-4-steps-turbo-GGUF - **License**: other - **Tasks**: text-to-image-synthesis - **Base model**: Viggle/Qwen-Image-2.1-viggle-turbo - **Tags**: gguf, text-to-image, image-to-image, comfyui, distilled, turbo, 4-step, qwen, qwen-image - **Downloads**: 590 - **Stars**: 2 - **Last updated**: 2026-09-23 Source: https://www.modelscope.cn/models/Abiray/Qwen-Image-2.1-viggle-4-steps-turbo-GGUF --- # Qwen-Image-2.1 Viggle Turbo GGUF (4-Step Distilled) This repository contains optimized community **GGUF quantizations** of [Viggle/Qwen-Image-2.1-viggle-turbo](https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo), an official 4-step distilled student model of Qwen-Image-2.1 trained by Viggle using **Distribution Matching Distillation (DMD)**. These files were converted directly from the official full transformer weights (`transformer/diffusion_pytorch_model.safetensors`), allowing you to run near-instant 4-step image generation and image editing natively inside **ComfyUI** without needing to stack external LoRAs. --- ## ⚡ Critical Generation Rules Because this model is distilled with DMD, **it requires specific inference parameters**. Using standard diffusion parameters will degrade image quality or cause visual burn artifacts: | Parameter | Value | Note | |---|---|---| | **Steps** | **`4 or 8 steps`** | Do not increase to 25 or 40. | | **CFG (Guidance Scale)** | **`1.0`** | Must remain at 1.0 (trained without classifier-free guidance). | | **Negative Prompt** | **Empty (Blank)** | Leave blank; CFG 1.0 ignores negative prompts. | | **Sampler** | **`euler`** | Standard Euler sampler. | | **Scheduler** | **`simple`** (or FlowMatch Euler) | Matches the distilled flow-matching trajectory. | | **Denoise** | **`1.0`** | For standard text-to-image generation. | --- ## 📦 Quantization Ladder All quants preserve high-precision **FP32** attention normalization layers (`norm_k`, `norm_q`, `text_norm`) to eliminate NaN errors and color banding, alongside native **BF16** patch input embeddings (`img_in.weight`). The `_M` tiers retain 6-bit precision (`Q6_K`) on critical attention value projections (`attn.to_v.weight`) and native **BF16** on text conditioning input layers (`txt_in`). | Filename | Quant Type | File Size | Recommended VRAM | Description | |---|---|---|---|---| | **`qwen_image_2.1_turbo_Q8_0.gguf`** | `Q8_0` | **7.59 GB** | 16 GB+ | Near-lossless 8-bit master quant. | | **`qwen_image_2.1_turbo_Q6_K.gguf`** | `Q6_K` | **5.88 GB** | 12 GB - 16 GB | High-detail tier with 6-bit attention. | | **`qwen_image_2.1_turbo_Q5_K_M.gguf`** | `Q5_K_M` | **5.01 GB** | 10 GB - 12 GB | Best quality-to-size balance (6-bit `to_v` heads). | | **`qwen_image_2.1_turbo_Q4_K_M.gguf`** | `Q4_K_M` | **4.19 GB** | 8 GB - 12 GB | **Recommended for most GPUs (RTX 3060/4060, Tesla T4).** | | **`qwen_image_2.1_turbo_Q4_K_S.gguf`** | `Q4_K_S` | **4.06 GB** | 8 GB | Slightly smaller 4-bit footprint. | | **`qwen_image_2.1_turbo_Q3_K_M.gguf`** | `Q3_K_M` | **3.19 GB** | 6 GB - 8 GB | Ultra-low VRAM tier. | --- ## 🛠️ Usage in ComfyUI ### 1. Requirements Ensure your ComfyUI and custom nodes are updated: * **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)** (Updated to latest) * **[ComfyUI-GGUF](https://github.com/city96/ComfyUI-GGUF)** (Updated to latest for `qwen_image` architecture recognition) * **Package Requirement:** `pip install -U "gguf>=0.13.0"` > **Troubleshooting Note:** If you encounter `ValueError: This model is not currently supported - (Unknown model architecture!)`, update your installation via `git pull` in both your `ComfyUI` root and `ComfyUI/custom_nodes/ComfyUI-GGUF` directories, then restart the server. ### 2. Model Placement Place your downloaded models into the standard ComfyUI directories: ComfyUI/ ├── models/ │ ├── diffusion_models/ │ │ └── qwen_image_2.1_turbo_Q4_K_M.gguf │ ├── text_encoders/ │ │ └── qwen3vl_8b_int8_convrot.safetensors (or fp8_scaled / w4a8) │ └── vae/ │ └── qwen_image_2.1_vae_bf16.safetensors ---- Supporting text encoder and VAE weights can be downloaded from Comfy-Org: * **Text Encoders:** [Comfy-Org/Qwen-Image-2.1 Text Encoders](https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/text_encoders) * **VAE:** [Comfy-Org/Qwen-Image-2.1 VAE](https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/vae/qwen_image_2.1_vae_bf16.safetensors) ### 3. Workflow Setup 1. Use **`UnetLoaderGGUF`** to load the selected Turbo GGUF file. 2. Load the standard Qwen3-VL text encoder (`CLIPLoader` set to type `qwen_image`) and VAE (`VAELoader`). 3. Connect conditioning via **`TextEncodeQwenImage21`**. 4. Set your **`KSampler`** to: * **`steps`**: `4` * **`cfg`**: `1.0` * **`sampler_name`**: `euler` * **`scheduler`**: `simple` * **`denoise`**: `1.0` 5. Leave the negative conditioning disconnected or prompt empty. --- ## 🎨 Image Editing Instructions Qwen-Image-2.1 natively handles instruction-based image editing directly: 1. Load your reference image using a standard **`Load Image`** node. 2. Connect the image output into the **`image_1`** input slot on the **`TextEncodeQwenImage21`** node. 3. Connect the VAE into the **`vae`** input slot on **`TextEncodeQwenImage21`**. 4. Enter your natural language editing instruction into the text prompt field (e.g., *"Change the background to a rainy cyberpunk street at night, keeping the woman's face and pose unchanged"*). 5. Run the sampler at `steps: 4` and `cfg: 1.0`. --- ## ⚠️ Notes & Limitations * **Distillation Preview:** This checkpoint is derived from Viggle's v0.1 preview release. It is optimized for fast Text-to-Image generation and rapid drafting at 1024×1024 resolution. * **Complex Multi-Reference Edits:** Complicated compositions or intricate text rendering may exhibit more variance than the full 40-step teacher model. For maximum typography fidelity, run the standard non-distilled Qwen-Image-2.1 models. --- ## 📜 Credits & License **Viggle** — Viggle/Qwen-Image-2.1-viggle-turbo * **Base Model:** Developed by [Qwen Team (Alibaba Cloud)](https://huggingface.co/Qwen) and distilled by [Viggle AI](https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo). * **GGUF Conversion:** Quantized by [Abiray](https://huggingface.co/Abiray) using `city96/ComfyUI-GGUF` and `llama.cpp`. * **License:** Derivative work distributed under the **Qwen Research License Agreement** (Non-commercial research and evaluation purposes).