Qwen-Image 2.1 Mixed Balanced
A ComfyUI-native mixed-precision quantization of Qwen/Qwen-Image-2.1, created with NidAll/comfyui-native-quantizer.
The goal is simple: keep Qwen-Image 2.1's image quality and typography while making the diffusion transformer much smaller and friendlier to low-VRAM GPUs.
Size
| Checkpoint | Size |
|---|---|
| BF16 source | 13.25 GiB |
| Mixed Balanced | 3.91 GiB |
| Reduction | 70.5% |
The conversion quantized 192 transformer linears, kept 73 tensors at source precision, and passed 65 validation checks with 0 warnings and 0 failures.
Quantization
The checkpoint was generated with:
comfyquant-cuda qwen_image_2.1_bf16.safetensors \
--format mixed \
--profile balanced \
--experimental \
--target-runtime cuda \
--output qwen_image_2.1_mixed_balanced.safetensors
Large attention and MLP linears are quantized, while sensitive paths such as inputs, timestep conditioning, modulation, attention norms, final normalization, and output projection remain at source precision.
Conversion summary:
architecture : qwen_image21
format : mixed / balanced
quantized : 192 layers
kept : 73 tensors
mean error : 0.0731
compression : 0.563 bytes/param
validation : 65 passed, 0 warnings, 0 failed
Example generations
Painterly composition
Typography-heavy comic
The comic is a useful stress test because it combines long text, multiple panels, small labels, repeated characters, and dense composition.
ComfyUI
Place the checkpoint in:
ComfyUI/models/diffusion_models/
Then use it with a normal Qwen-Image 2.1 workflow.
Recommended companion components for low-VRAM setups:
- Qwen3-VL 8B W4A8 text encoder
- Qwen-Image 2.1 BF16 VAE
The text encoder and VAE are not included in this repository.
Quality notes
Initial testing shows very strong retention compared with the official INT8 ConvRot model, including typography-heavy generations.
Outputs are not expected to be pixel-identical to BF16 or INT8 at the same seed. Small numerical differences can change the diffusion trajectory while preserving overall prompt adherence and visual quality.
This checkpoint is experimental, so BF16 or the official INT8 model remain useful references for maximum reproducibility.
Credits
License
This is an independent quantized derivative of Qwen-Image 2.1. Use remains subject to the applicable upstream Qwen-Image 2.1 license terms.
- Downloads last month
- 491
Model tree for NidAll/Qwen-Image-2.1-Mixed-Balanced
Base model
Qwen/Qwen-Image-2.1