Qwen-Image 2.1 Uncensored (GGUF) — ComfyUI on Google Colab

ComfyUI setup for abenzerps/Qwen-Image-2.1-Uncensored-GGUF: the uncensored (UC) fine-tune of Qwen/Qwen-Image-2.1 (the author's LoRA merged into the upstream weights) as GGUF quants, an unquantized BF16 GGUF and fp8 / int8 safetensors. The plain base model stays available (--model base-*). Text-to-image (native 2K, N images per prompt), image edit via reference images, LoRAs.

Click the badge to open the notebook straight from this repo, or see COLAB.md for the cell-by-cell guide.

Layout

NewGwen2.1/
├── Qwen_Image_2.1_GGUF_ComfyUI.ipynb   # Colab notebook (runs the scripts below)
├── COLAB.md                            # same steps as prose
├── setup_colab.sh                      # ComfyUI + leejet/ComfyUI-GGUF at pinned commits + deps
├── download_models.py                  # diffusion model (UC / base, GGUF or safetensors) + text encoder + VAE + LoRAs by URL
├── launch_comfyui.py                   # start server, print Colab proxy / cloudflared URL
├── app_gradio.py                       # Gradio UI on *.gradio.live: t2i + edit tabs, N images, Send to Edit, LoRAs, host status
├── generate.py                         # headless t2i (run) / edit (edit) via the ComfyUI API; --count N, --model, --lora
├── colab_requirements.txt
├── packages.txt                        # apt packages (none needed at the moment)
└── workflows/
    ├── qwen_image_2.1_gguf_t2i.json          # text to image — drag into ComfyUI
    ├── qwen_image_2.1_gguf_t2i_api.json      # … API format, used by generate.run
    ├── qwen_image_2.1_gguf_edit.json         # image edit (LoadImage x2 → TextEncodeQwenImage21) — drag into ComfyUI
    ├── qwen_image_2.1_gguf_edit_api.json     # … API format, used by generate.edit
    ├── qwen_image_2.1_safetensors_t2i.json   # same graphs with the stock UNETLoader for the
    ├── qwen_image_2.1_safetensors_edit.json  # … fp8 / int8 / base-bf16 *.safetensors
    └── build_workflows.py                    # regenerates all of them (e.g. --quant Q8_0)

Quick start (Colab)

!git clone https://github.com/AndreRese/NewGwen2.1 qwen21
%cd qwen21
!bash setup_colab.sh
!python download_models.py --model Q4_K_M --text-encoder int8     # or --model BF16 for unquantized
!python app_gradio.py     # ComfyUI canvas (Colab proxy) + Gradio UI (*.gradio.live)

Model files

*.gguf loads through Unet Loader (GGUF), *.safetensors through the stock Load Diffusion Model (UNETLoader); generate.py / the Gradio app pick the loader from the extension. The repo also has Q5_K_M / Q6_K; they are left out of the menu on purpose.

LoRAs

Only Qwen-Image 2.1 LoRAs work (1.x / 2.0 ones load with lora key not loaded in the log and change nothing). Put .safetensors files in ComfyUI/models/loras, or download them by URL:

!python download_models.py --lora-only --lora https://huggingface.co/<user>/<repo>/blob/main/<file>.safetensors
!python download_models.py --lora-only --lora "https://civitai.com/api/download/models/<id>" --lora-token <API key>

The Gradio app has 3 LoRA slots (applied to both tabs) and the same download box; generate.py --lora file.safetensors:0.8 (repeatable) and run(..., loras=[(name, strength)]) chain LoraLoaderModelOnly nodes after the diffusion loader. The canvas workflows carry one bypassed LoRA node (Ctrl+B to enable). Only .safetensors is accepted: .ckpt / .pt can run code.

Pinned versions

setup_colab.sh checks out ComfyUI b0f4b7b2 (2026-09-21, has the Qwen-Image 2.1 nodes) and leejet/ComfyUI-GGUF edd981b1. The stock city96/ComfyUI-GGUF does not know the Qwen-Image 2.1 architecture; the leejet fork (author of stable-diffusion.cpp, which produced these quants) added it on 2026-09-20. UPDATE_TO_LATEST=1 bash setup_colab.sh uses ComfyUI master + the newest loader instead; bump the pins in the script once a newer pair is tested.

History: the Q8_0 norm-weight crash (fixed 2026-09-21)

The first qwen-image-2.1-Q8_0.gguf had its 1D norm_q/norm_k weights stored as Q8_0 blocks, which crashed KSampler with Expected weight … shape [136] and normalized_shape = [128] (discussion #4). Fixed on both sides: the file was re-uploaded and the pinned loader dequantizes every quantized 1D tensor.

Workflow

UnetLoaderGGUF (or the stock UNETLoader for .safetensors) → KSampler (euler / simple, 25 steps, cfg 1.0) with CLIPLoader(type=qwen_image) → TextEncodeQwenImage21 for conditioning, EmptyLatentImage 1024² (or 2048² for native 2K) and VAEDecode → SaveImage. This is Comfy-Org's official image_qwen_image_2_1_t2i template, flattened out of its subgraph, with the loader swapped for GGUF.

Image edit (image_qwen_image_2_1_image_edit template, same treatment): LoadImage → TextEncodeQwenImage21 (images.image_1…16 + VAE); image_1 is the edit target and its latent output sets the canvas; QwenImage21Cache (KV cache device / int8 / int4) between the loader and the sampler. Prompt with <image1>, <image2>, …