license: other
license_name: qwen-research-license
base_model: Qwen/Qwen-Image-2.1
tags:
- gguf
- qwen-image
- text-to-image
- image-editing
- comfyui
- uncensored
Qwen-Image-2.1-Uncensored-step90000-GGUF
GGUF quantizations of the merged LoRA checkpoint Qwen-Image-2.1-Uncensored-step90000.
Source: qwen_image_2.1_step90000_merged_bf16.safetensors (14.23 GB, 297 tensors, 32 blocks).
Architecture: qwen_image21, key names kept in diffusers form (img_in.weight, transformer_blocks.N.attn.to_q.weight, img_mlp.gate_layer/proj, ...).
Files
| File | Size |
|---|---|
qwen-image-2.1-s90000-BF16.gguf |
14.23 GB |
qwen-image-2.1-s90000-Q8_0.gguf |
7.59 GB |
qwen-image-2.1-s90000-Q6_K.gguf |
5.88 GB |
qwen-image-2.1-s90000-Q5_K_M.gguf |
5.22 GB |
qwen-image-2.1-s90000-Q4_K_M.gguf |
4.60 GB |
qwen-image-2.1-s90000-Q4_0.gguf |
4.15 GB |
Recommended: Q4_K_M.
Quantization rules
| Variant | Rule |
|---|---|
| BF16 | original |
| Q8_0 | all non-protected tensors Q8_0 |
| Q6_K | all non-protected tensors Q6_K |
| Q5_K_M | attn.to_v / img_mlp.out Q6_K, rest Q5_K |
| Q4_K_M | attn.to_v / img_mlp.out Q6_K, rest Q4_K |
| Q4_0 | img_mlp.out Q4_1, rest Q4_0 |
Protected layers kept BF16 in every variant: img_in.*, txt_in.*, all attn.norm_q/norm_k (68 tensors).
Verification
Full 297-tensor dequantized comparison against the bf16 source (global relative RMSE):
| Variant | RMSE |
|---|---|
| BF16 | bit-identical |
| Q8_0 | 0.565 % |
| Q6_K | 1.808 % |
| Q5_K_M | 3.167 % |
| Q4_K_M | 6.062 % |
| Q4_0 | 8.321 % |
Protected layers are bit-identical to the source in both checkpoints. SHA256SUMS is included.
ComfyUI
Requires ComfyUI with Qwen-Image-2.1 support and ComfyUI-GGUF (arch list includes qwen_image21).
ComfyUI/models/
├── diffusion_models/ qwen-image-2.1-s90000-Q4_K_M.gguf
├── text_encoders/ qwen3vl_8b_bf16.safetensors
└── vae/ qwen_image_2.1_vae_bf16.safetensors
Nodes: Unet Loader (GGUF) + CLIPLoader (type=qwen_image) + TextEncodeQwenImage21.
Sampler: euler + simple, 25–40 steps, cfg 1.0, denoise 1.0. Resolution 1024×1024 and up (multiples of 32).
Companion files: chatrlabs/Qwen-Image-2.1-Uncensored-companion.
License
Qwen Research License Agreement — see LICENSE. Check the terms before commercial use or redistribution.