Sponsored by
Vmake Labs Video Enhancer
-Vmake Labs Video Enhancer uses AI to improve video quality, restore details, reduce blur, and upscale videos to 8K.
Z-Image-Turbo
by Tongyi-MAI, packaged as a single
CMF
file: the 6B image transformer, the Qwen3-4B
text encoder and the VAE together, with the generation recipe stored inside. It runs
with
cortiq
, a Rust engine with no Python: NVIDIA GPUs through Vulkan
(tensor cores), Apple silicon through Metal, and a CPU fallback everywhere.
A cat on a windowsill at sunset, photorealistic
A fisherman's hands tying a rope
A "CORTIQ" neon sign on a brick wall
An aerial view of a river in autumn
1024×1024, seed 7, the file's default settings, generated by cortiq on the fast path (8 steps); click for the full PNG.
Try it in the browser:
🤗 Space
(free CPU tier: a 256² Turbo image takes about 10 minutes; the tables below are what a GPU or a Mac does).
cargo install cortiq-cli # 0.7.5 or later; prebuilt binaries: github.com/infosave2007/cmf/releases
hf download infosave/Z-Image-Turbo-cmf z-image-turbo.cmf --local-dir .
cortiq imagine z-image-turbo.cmf --prompt "A cat sitting on a windowsill at sunset, photorealistic"
No flags needed: the file carries the Turbo recipe (1024×1024, 8 steps, no CFG). Set
XDG_RUNTIME_DIR=/tmp
on a headless Linux box.
File
file
size
inside
z-image-turbo.cmf
10.5 GB
DiT 8-bit (
q8_2f
), Qwen3-4B text encoder 8-bit, VAE
sha256
c0aaffaf887c595431cddfb0dce369d04eba23dac7a69ca162a051486564f1f8
(
z-image-turbo.cmf.sha256
is next to it). 8-bit was chosen by measurement: the 4-bit
transformer loses 3–15 dB of image PSNR, the 8-bit one is comparable to bf16 inference.
Performance
RTX 3090, Vulkan, one image per process, no flags (cortiq 0.7.5):
size
time to image
of which DiT steps
VAE
512×512
4.6 s
2.1 s (8 steps)
0.13 s
1024×1024
11.4 s
8.3 s (8 steps)
0.50 s
Mac mini M4 (24 GB), Metal, no flags:
size
time to image
of which DiT steps
VAE
512×512
30 s
28.5 s (8 steps)
1.0 s
1024×1024
160 s
155 s (8 steps)
4.8 s
About 8 GB of memory. The GEMMs run at 91–94 % of the M4 GPU's matrix peak; attention is the remaining gap.
RTX 3090: peak VRAM 14–15 GB (16 GB cards fit). Per image, the DiT runs at the speed of diffusers
bf16 on the same card and the whole process starts about four times faster.
Options
option
meaning
--width
,
--height
multiples of 16 (default 1024×1024)
--steps
,
--seed
,
--num-images N
image i uses seed+i
--cfg G
,
--negative-prompt
guidance scale and negative prompt (Turbo: 0, no CFG)
--cfg-normalization [C]
,
--cfg-truncation T
the diffusers CFG extras
--shift
,
--max-sequence-length
scheduler shift, prompt token cap (512)
--out
.png
(default
out.png
),
.jpg
,
.ppm
CMF_ZIMAGE_PROF=1
prints stage times.
CMF_GPU=0
runs everything on the CPU (minutes per image).
Turbo is distilled: guidance stays at 0 and 8 steps are the recipe; negative prompts have no effect on it — use
Z-Image
for CFG and negative prompting.
sha256sum -c z-image-turbo.cmf.sha256
cortiq info z-image-turbo.cmf
Документация на русском
Z-Image-Turbo
от Tongyi-MAI одним файлом
CMF
: трансформер 6B, текстовый энкодер Qwen3-4B и VAE
вместе, рецепт генерации записан в файл. Запускается движком
cortiq
на Rust без Python:
NVIDIA через Vulkan (тензорные ядра), Apple silicon через Metal, везде есть путь на CPU.
cargo install cortiq-cli # 0.7.5 или новее; готовые бинарники: github.com/infosave2007/cmf/releases
hf download infosave/Z-Image-Turbo-cmf z-image-turbo.cmf --local-dir .
cortiq imagine z-image-turbo.cmf --prompt "Кот на подоконнике на закате, фотореализм"
Флаги не нужны: в файле записан рецепт Turbo (1024×1024, 8 шагов, без CFG). На Linux без дисплея задайте
XDG_RUNTIME_DIR=/tmp
.
Файл:
z-image-turbo.cmf
, 10.5 ГБ — DiT 8 бит (
q8_2f
), энкодер Qwen3-4B 8 бит, VAE;
sha256
c0aaffaf887c5954…
(полностью в
z-image-turbo.cmf.sha256
). 8 бит выбраны по замеру: 4-битный трансформер
теряет 3–15 дБ PSNR картинки, 8-битный сравним с выводом в bf16.
Скорость
(RTX 3090, Vulkan, одна картинка на процесс, без флагов, cortiq 0.7.5):
размер
картинка целиком
из них шаги DiT
VAE
512×512
4.6 с
2.1 с (8 шагов)
0.13 с
1024×1024
11.4 с
8.3 с (8 шагов)
0.50 с
Mac mini M4 (24 ГБ), Metal, без флагов: 512² за 30 с (шаги DiT 28.5 с, VAE 1.0 с),
1024² за 160 с (шаги 155 с, VAE 4.8 с); память около 8 ГБ. GEMM идут на 91–94 % матричного
пика GPU M4; остаток — внимание.
RTX 3090: пик видеопамяти 14–15 ГБ (хватает карты на 16 ГБ). DiT идёт со скоростью diffusers bf16 на той же
карте, а старт процесса примерно в четыре раза быстрее.
Опции:
--width
/
--height
(кратны 16, по умолчанию 1024×1024),
--steps
,
--seed
,
--num-images N
(картинка i берёт seed+i),
--cfg G
и
--negative-prompt
(Turbo: 0, без CFG),
--cfg-normalization [C]
,
--cfg-truncation T
,
--shift
,
--max-sequence-length
(512),
--out
(
.png
по умолчанию,
.jpg
,
.ppm
).
CMF_ZIMAGE_PROF=1
печатает время стадий,
CMF_GPU=0
считает всё на CPU (минуты на картинку).
Turbo дистиллирована: guidance 0 и 8 шагов — это её рецепт, негативный промпт на неё не действует; для CFG и негативных промптов есть
Z-Image
.
Примеры:
1024², seed 7, настройки по умолчанию —
samples/
.
Проверка:
sha256sum -c z-image-turbo.cmf.sha256
,
cortiq info z-image-turbo.cmf
.
Z-Image-Turbo-cmf huggingface.co is an AI model on huggingface.co that provides Z-Image-Turbo-cmf's model effect (), which can be used instantly with this infosave Z-Image-Turbo-cmf model. huggingface.co supports a free trial of the Z-Image-Turbo-cmf model, and also provides paid use of the Z-Image-Turbo-cmf. Support call Z-Image-Turbo-cmf model through api, including Node.js, Python, http.
Z-Image-Turbo-cmf huggingface.co is an online trial and call api platform, which integrates Z-Image-Turbo-cmf's modeling effects, including api services, and provides a free online trial of Z-Image-Turbo-cmf, you can try Z-Image-Turbo-cmf online for free by clicking the link below.
infosave Z-Image-Turbo-cmf online free url in huggingface.co:
Z-Image-Turbo-cmf is an open source model from GitHub that offers a free installation service, and any user can find Z-Image-Turbo-cmf on GitHub to install. At the same time, huggingface.co provides the effect of Z-Image-Turbo-cmf install, users can directly use Z-Image-Turbo-cmf installed effect in huggingface.co for debugging and trial. It also supports api for free installation.