Running Uncensored Qwen-Image-2.1 on Laptops: Q4_K_M (4.68GB) and Text Encoder Swap in ComfyUI

Image source: https://huggingface.co/pottokao/Qwen-Image-2.1-Text-Encoder-Heretic-GGUF

AI creator and engineer 三个三 (@3three_AI) has shared a practical local workflow for running an uncensored setup of Alibaba's Qwen-Image-2.1 on laptops and mainstream 8GB VRAM consumer GPUs without prompt refusals, combining a 4.68GB Q4_K_M quantized diffusion backbone with a community-modified Heretic text encoder inside ComfyUI.

Qwen-Image-2.1 introduced a unified architecture that handles text-to-image generation, multi-image reference editing, and native alpha transparency in a single model. However, strict default alignment layers in the official release frequently triggered false-positive prompt refusals during creative workflows. Because the refusal behavior is concentrated within the multimodal text encoder rather than the visual diffusion backbone, creators can resolve this bottleneck by decoupling the pipeline—quantizing the 7B DiT backbone for low memory consumption while swapping the text encoder for an uncensored community build.

Architecture Breakdown: 7B DiT Q4_K_M and Decoupled Text Encoding

Running foundation image models on consumer hardware requires managing video memory across multiple concurrent model components.

Qwen-Image-2.1 combines three distinct neural modules: a 7B-parameter 32-layer single-stream Diffusion Transformer (DiT) visual backbone, a Qwen3-VL 8B text encoder, and a 64-channel RGBA autoencoder (VAE) with 16x spatial compression. In the official BF16 release, the 8B text encoder spans approximately 17.5GB across four shards, making it significantly larger than the 7B visual backbone itself (~14.2GB across two shards).

• Diffusion Backbone Quantization: The 7B DiT backbone is quantized to Q4_K_M GGUF (~4.68GB) using the city96 ComfyUI-GGUF format. This reduction lowers the base footprint sufficiently to fit within the memory headroom of mainstream 8GB GPUs such as mobile RTX 3060 or desktop RTX 4060 cards.
• Targeted Text Encoder Swap: Rather than retraining the visual backbone or applying aggressive quantization that degrades prompt nuance, the refusal-heavy Qwen3-VL 8B text encoder is replaced with Qwen-Image-2.1-Text-Encoder-Heretic-GGUF, curated by developer pottokao.
• Refusal Drop to 5%: Swapping to the Heretic encoder drops the system refusal metric from approximately 100 down to 5. Crucially, standard prompt following, descriptive vocabulary, and composition comprehension remain intact without degrading general generation quality.

Step-by-Step ComfyUI Installation and Node Pipeline Configuration

Deploying this configuration locally in ComfyUI requires placing the respective weight files into standard directories and wiring the appropriate GGUF loader nodes.

1. File Placement and Directory Setup

Download the verified model files and place them into the corresponding ComfyUI subdirectories:

1. Diffusion Model Backbone: Download the Qwen-Image-2.1 7B DiT Q4_K_M GGUF file (qwen_image_2.1_Q4_K_M.gguf, ~4.68GB) and place it into ComfyUI/models/diffusion_models/ (or models/unet/).
2. Text Encoder: Download Qwen-Image-2.1-Text-Encoder-Heretic-GGUF from Hugging Face and place it into ComfyUI/models/text_encoders/ (or models/clip/).
3. RGBA VAE: Ensure the official Qwen-Image-2.1 64-channel RGBA VAE is placed in ComfyUI/models/vae/.

2. Node Graph Connections

Configure the node workflow using the ComfyUI-GGUF extension nodes:

• Add the UnetLoaderGGUF node provided by the city96 ComfyUI-GGUF custom extension and select the qwen_image_2.1_Q4_K_M.gguf file.
• Connect the text encoder loader pointing to Qwen-Image-2.1-Text-Encoder-Heretic-GGUF to supply text conditioning to the sampler.
• Wire the official 64-channel RGBA VAE to the VAE decode node. Because Qwen-Image-2.1 natively outputs RGBA channels, sticker graphics, transparent assets, and cutouts render directly with clean alpha channels, completely bypassing manual post-process masking or background segmentation tools.

Multi-Image Reference Editing and 8GB VRAM Hardware Caveats

Qwen-Image-2.1 natively supports multi-image editing workflows, including inpainting via localized brush or circle annotations, as well as conditioning on up to 10 visual reference images simultaneously.

While these editing features are fully integrated into the architecture, operating near an 8GB hardware ceiling requires careful memory discipline:

• Reference Image Memory Scaling: Conditioning on multiple reference images (up to 10 images) causes cross-attention maps and latent tensor buffers to expand rapidly, resulting in substantial VRAM spikes during the initial sampling passes.
• Baseline Generation Sizing: For systems equipped with 8GB VRAM, creators should begin with single 1024x1024 text-to-image generations to establish a baseline before attempting multi-reference compositions or higher output dimensions.
• Operational Memory Hygiene: To prevent out-of-memory (OOM) allocation failures during VAE decoding, close GPU-accelerated browser tabs and background applications before executing generations, and enable VAE tiling options in ComfyUI when decoding larger image resolutions.

Original source

• X (formerly Twitter) @3three_AI Original Post: Qwen-Image-2.1 Uncensored Local Setup Guide
• Hugging Face Repository: Qwen-Image-2.1-Text-Encoder-Heretic-GGUF by pottokao