license: other
license_name: qwen-research
base_model:
- Qwen/Qwen-Image-2.1
base_model_relation: quantized
pipeline_tag: text-to-image
library_name: cortiq
tags: - cmf
- cortiq
- qwen-image
- text-to-image
- image-generation
- diffusion
- quantized
- vulkan
- metal
- cpu
- local-inference
- uncensored
language: - en
Qwen-Image-2.1-Uncensored-CMF
A verified, single-file CMF build of the Qwen-Image-2.1 pipeline for local image generation with Cortiq.
12.745 GB · 15.59B parameters · one file · CPU, native Metal, and Vulkan runtime paths
The file contains the DiT, text/vision encoders, VAE, tokenizer, and recipe. Its production layout is retained from the validated CMF base: q4tp DiT + q8_2f encoders + f16 VAE. This is a runtime-oriented layout, not a raw BF16 dump.
Get started
Download qwen-image-2.1-uncensored.cmf and its matchingqwen-image-2.1-uncensored.cmf.sha256 file from the Files tab, then verify
the download:
shasum -a 256 -c qwen-image-2.1-uncensored.cmf.sha256
To download both files from a terminal instead, use the Hugging Face CLI:
hf download infosave/Qwen-Image-2.1-Uncensored-CMF \
qwen-image-2.1-uncensored.cmf qwen-image-2.1-uncensored.cmf.sha256 \
--local-dir ./image-2.1-uc
| File | Value |
|---|---|
| Artifact | qwen-image-2.1-uncensored.cmf |
| Exact size | 12,745,219,006 bytes (11.870 GiB) |
| SHA-256 | 44317a2e7ebe4a520ba2e701ff099f8faaeafe065972395c3ac53ae32f99aada |
| Tested runtime | cortiq 0.8.2 |
Generate an image
cargo install cortiq-cli --locked
cortiq imagine qwen-image-2.1-uncensored.cmf \
--prompt "A red fox in fresh snow at golden hour" \
--out fox.png
The artifact was tested with cortiq 0.8.2. The embedded recipe defaults to 1024×1024 and 40 steps. For a fast installation check:
cortiq verify qwen-image-2.1-uncensored.cmf
cortiq imagine qwen-image-2.1-uncensored.cmf \
--prompt "A red fox in fresh snow at golden hour" \
--width 512 --height 512 --steps 4 --seed 42 --out smoke.png
Examples
1024×1024 · 40 steps · cfg=1 · seed=42
Prompt: A red fox in fresh snow at golden hour
1024×1024 · 40 steps · cfg=1 · seed=271828
Prompt: A futuristic botanical observatory built into a cliff above a calm ocean at sunrise, glass domes, lush bioluminescent plants, cinematic wide-angle architectural visualization
Both images were generated from this exact artifact on the RTX 4090 Vulkan f16-fallback path.
Hardware paths
| Platform | Command | Status |
|---|---|---|
| CPU | CMF_GPU=0 cortiq imagine … |
Tested. Use CMF_THREADS to tune the host. |
| NVIDIA / Vulkan | XDG_RUNTIME_DIR=/tmp WGPU_BACKEND=vulkan CMF_GPU=1 cortiq imagine … |
Tested on RTX 4090. |
| Apple silicon / Metal | CMF_GPU=1 CMF_QI21_METAL_PROF=1 cortiq imagine … |
Native resident path is retained by this layout. Measure on the target Mac before publishing a Metal number. |
For headless RunPod-style containers, install the Vulkan userspace once before testing:
apt-get update && apt-get install -y libglvnd0 libgl1 libegl1 libvulkan1 vulkan-tools
XDG_RUNTIME_DIR=/tmp vulkaninfo --summary | grep deviceName
Measured Vulkan result
This is a run of this exact artifact, not a number copied from another Qwen model:
| Host | Workload | Path | Result |
|---|---|---|---|
| RTX 4090, driver 570.195.03, Vulkan | 1024×1024, 40 steps, cfg=1, seed=42 | Vulkan f16 fallback | 29.13 s end to end; 0.527 s median DiT step |
The run held 13.96 GB of DiT planes on the device; the 1024² VAE decode reported 4.8 GB activations and 0.81 s decode time. The driver exposed no float8 matrix extension (fp8 mma false), so this is honestly labeled f16 fallback, not FP8.
CPU tuning on AMD EPYC 7642 (96 logical CPUs), using 256×256 / 2 steps / the same fixed prompt and seed:
CMF_THREADS |
Total | Median DiT step |
|---|---|---|
| 48 | 86.70 s | 32.720 s |
| 32 | 66.05 s | 24.317 s |
| 12 | 43.54 s | 15.806 s |
For this host and workload, 12 threads beat the default-like 32-thread configuration by about 34% end to end. This is a tuning result, not a universal CPU claim. Repeating the 12-thread run produced the identical PNG SHA-256.
See BENCHMARKS.md for the reproducible protocol and limitations.
What “UC” means
UC identifies the requested source input:abenzerps/Qwen-Image-2.1-Uncensored-GGUF.
Its publisher states that its GGUF release has no built-in safety checker or content filter.
That is an attributed upstream statement, not a guarantee about every prompt or output of this CMF artifact. Applications and operators remain responsible for their own safeguards, policies, and legal compliance.
Conversion provenance
- Validated CMF pipeline base SHA-256:
b94309e4ed716c7a224c3dc88d583d3095d6e34edd76a76c01b04b9123a40bce - Source DiT:
qwen-image-2.1-UC-BF16.gguf - Source DiT SHA-256:
f151c683a8aed4b310777017ebbbe3f2180f1180f7867115171adb7d50b0762a - Converter integration:
34d33dc8db5a085210794992940f67f3ec5accf4 - Assembly mode: full-pipeline base validation; only DiT tensors are replaced and re-encoded once into the base slots. Encoders, VAE, tokenizer, scheduler, and non-DiT payloads are copied byte-for-byte.
License
This is a modified redistribution under the Qwen RESEARCH LICENSE AGREEMENT. It is for non-commercial research use; commercial use requires a separate license from the Qwen licensor. See NOTICE.
Built with Qwen Image 2.1.