infosave/Image-2.1-cmf

Image-2.1 — CMF

Qwen-Image-2.1 packaged as a single CMF file: the 7B image transformer, the Qwen3-VL-8B text encoder with its vision tower, and the RGBA VAE together, with the generation recipe stored inside. It runs with cortiq, a Rust engine with no Python: Apple silicon through Metal, NVIDIA GPUs through Vulkan, and a CPU fallback everywhere. Text-to-image, native transparency (RGBA), and generation from condition images. Built with Qwen.

A neon shop sign that reads "CORTIQ", rainy night, reflections on wet pavement	RGBA: a cute cartoon dragon sticker, transparent background
A red fox sitting in fresh snow at golden hour, photorealistic	A vintage travel poster of Lake Baikal, the title "BAIKAL"
--image (left, the fox above) + "Change the background to a sunset beach, keep the fox unchanged" (right)

_{1024×1024, 40 steps, default settings, generated by cortiq (Vulkan, RTX PRO 4000); click for the full PNG.}

English · Русский · 中文

Quick start

cargo install cortiq-cli          # 0.8.2 or later; prebuilt binaries: github.com/infosave2007/cmf/releases
hf download infosave/Image-2.1-cmf qwen-image-2.1.cmf qwen-image-2.1.cmf.sha256 --local-dir .
cortiq imagine qwen-image-2.1.cmf --prompt "A neon shop sign that reads \"CORTIQ\", rainy night, reflections on wet pavement"

No flags needed: the file carries the recipe (1024×1024, 40 steps, no CFG). Set XDG_RUNTIME_DIR=/tmp on a headless Linux box.

Transparent images — ask for them in the prompt, PNG keeps the alpha channel:

cortiq imagine qwen-image-2.1.cmf --out sticker.png --prompt "This is an RGBA image with transparency. A cute cartoon dragon sticker. The image has alpha channel and the background is transparent."

From condition images (repeat --image for several):

cortiq imagine qwen-image-2.1.cmf --image photo.png --prompt "Change the background to a sunset beach"

File

file	size	inside
qwen-image-2.1.cmf	12.7 GB	transformer 4-bit (q4tp), Qwen3-VL-8B text encoder and vision tower 8-bit (q8_2f), VAE f16

sha256 b94309e4ed716c7a224c3dc88d583d3095d6e34edd76a76c01b04b9123a40bce (qwen-image-2.1.cmf.sha256 is next to it). The text encoder is 8-bit by measurement: at 4 bits its output moves 40 % against the fp32 reference, at 8 bits 11 %. With bf16 weights the engine matches the diffusers fp32 reference to 1e-5.

Performance

RTX PRO 4000 Blackwell (24 GB), Vulkan, one image per process, default settings (40 steps):

task	size	time to image	DiT steps	prompt encode	VAE (load + decode)
text-to-image	512×512	15 s	9.6 s (0.24 s a step)	2.4 s	1.4 s
text-to-image	1024×1024	48 s	41.6 s (1.04 s a step)	2.2 s	1.8 s
text-to-image	2048×2048	246 s	238 s (6.0 s a step)	2.0 s	3.2 s
one 1024² condition image	1024×1024	122 s	47 s (1.18 s a step)	60 s	1.7 s

Peak VRAM 17.5 GB at 1024² (24 GB cards). With several GPUs the engine takes the one with tensor cores and the most memory; CMF_GPU_ADAPTER=<n> (see cortiq gpu) pins another.

Mac mini M4 (10-core GPU, 24 GB), Metal:

task	size	time to image	DiT steps	VAE (load + decode)
text-to-image	512×512	3.9 min	5.5 s a step	7.6 s
text-to-image	1024×1024	20 min	25–30 s a step	34 s

About 9 GB of memory. The GEMMs run at ≈ 83 % of the M4 GPU's matrix peak, so the time is the model's arithmetic; --steps 20 halves it. Measured with cortiq 0.8.2.

Options

option	meaning
--width, --height	multiples of 32 (default 1024×1024, up to 2048² and beyond on 24 GB; with --image, the last image's aspect)
--steps, --seed, --num-images N	image i uses seed + i
--image PATH	condition image, repeatable
--reference-size S	condition images are resized to S² area (default 1024)
--cfg G, --negative-prompt	true CFG when G > 1 and a negative prompt is given (the model samples without guidance)
--out	.png (default out.png, RGBA), .jpg, .ppm

CMF_QI21_PROF=1 prints stage times. CMF_GPU=0 runs everything on the CPU (very slow).

License

Qwen-Image-2.1 is released under the Qwen Research License: non-commercial use only (research and evaluation). This repository redistributes it modified — quantized and repackaged into CMF by infosave — under the same agreement; see LICENSE and NOTICE. Commercial use needs a separate license from the Qwen team. Built with Qwen.

Verify

sha256sum -c qwen-image-2.1.cmf.sha256    # macOS: shasum -a 256 -c
cortiq info qwen-image-2.1.cmf

---

Документация на русском

Qwen-Image-2.1 одним файлом CMF: трансформер 7B, текстовый энкодер Qwen3-VL-8B с vision-башней и RGBA-VAE вместе, рецепт генерации записан в файл. Запускается движком cortiq на Rust без Python: Apple silicon через Metal, NVIDIA через Vulkan, везде есть путь на CPU. Генерация по тексту, нативная прозрачность (RGBA), генерация по картинкам-условиям.

cargo install cortiq-cli          # 0.8.2 или новее; готовые бинарники: github.com/infosave2007/cmf/releases
hf download infosave/Image-2.1-cmf qwen-image-2.1.cmf qwen-image-2.1.cmf.sha256 --local-dir .
cortiq imagine qwen-image-2.1.cmf --prompt "Лиса сидит в свежем снегу, золотой час, фотореализм"
cortiq imagine qwen-image-2.1.cmf --image photo.png --prompt "Замени фон на закат на пляже"

Флаги не нужны: 1024×1024, 40 шагов, без CFG. Прозрачный фон — попросите в промпте, PNG сохраняет альфа-канал. На Linux без дисплея задайте XDG_RUNTIME_DIR=/tmp.

Файл: qwen-image-2.1.cmf, 12.7 ГБ — трансформер 4 бит (q4tp), энкодер и vision-башня 8 бит (q8_2f), VAE f16; sha256 в qwen-image-2.1.cmf.sha256.

Скорость (один процесс на картинку, настройки по умолчанию, 40 шагов):

устройство	512²	1024²	2048²	1024² с картинкой-условием
RTX PRO 4000 Blackwell, Vulkan	15 с	48 с (шаг 1.04 с)	246 с (шаг 6.0 с)	122 с (энкодер 60 с)
Mac mini M4 24 ГБ, Metal	3.9 мин	20 мин (шаг 25–30 с)	—	—

Пик видеопамяти 17.5 ГБ (карты на 24 ГБ); из нескольких GPU берётся карта с тензорными ядрами и большей памятью, CMF_GPU_ADAPTER=<n> выбирает другую. На Mac около 9 ГБ памяти, GEMM идут на ≈ 83 % матричного пика M4; --steps 20 вдвое быстрее.

Опции: --width/--height (кратны 32), --steps, --seed, --num-images N, --image (повторяемый), --reference-size S, --cfg G с --negative-prompt, --out (.png с альфой, .jpg, .ppm). CMF_QI21_PROF=1 печатает время стадий.

Лицензия: Qwen Research License — только некоммерческое использование; файлы изменены (квантованы и упакованы в CMF), см. LICENSE и NOTICE. Built with Qwen.

---

中文文档

Qwen-Image-2.1 打包为单个 CMF 文件：7B 图像 Transformer、Qwen3-VL-8B 文本编码器及其视觉塔、 RGBA VAE 合在一起，生成配方也写在文件里。由 Rust 引擎 cortiq 运行，无需 Python：Apple silicon 走 Metal，NVIDIA 走 Vulkan，任何机器都有 CPU 后备路径。支持文生图、原生透明（RGBA）和以图为条件的生成。

cargo install cortiq-cli          # 0.8.2 或更新
hf download infosave/Image-2.1-cmf qwen-image-2.1.cmf qwen-image-2.1.cmf.sha256 --local-dir .
cortiq imagine qwen-image-2.1.cmf --prompt "雨夜里写着 \"CORTIQ\" 的霓虹招牌，湿润路面上的倒影"

无需参数：1024×1024，40 步，无 CFG。透明背景请在提示词中说明，PNG 保留 alpha 通道。

文件： qwen-image-2.1.cmf，12.7 GB —— Transformer 4 位（q4tp）、文本编码器与视觉塔 8 位（q8_2f）、VAE f16。

性能（每进程一张图，默认设置，40 步）：

设备	512²	1024²	2048²	1024² 带条件图
RTX PRO 4000 Blackwell，Vulkan	15 s	48 s（每步 1.04 s）	246 s（每步 6.0 s）	122 s（编码 60 s）
Mac mini M4 24 GB，Metal	3.9 分钟	20 分钟（每步 25–30 s）	—	—

显存峰值 17.5 GB（24 GB 显卡）；多卡时自动选择带张量核心、显存最大的卡，CMF_GPU_ADAPTER=<n> 可指定。Mac 约占 9 GB 内存，--steps 20 可减半耗时。

许可： Qwen Research License —— 仅限非商业用途；文件经过修改（量化并打包为 CMF），见 LICENSE 与 NOTICE。Built with Qwen。