CivArchive
    ZIT-Prism - v1.7
    NSFW
    Preview 130945262
    Preview 130945310
    Preview 130945395
    Preview 130945412
    Preview 130945479
    Preview 130945514
    Preview 130945575
    Preview 130945582
    Preview 130945622
    Preview 130945698

    🌟 ZIT-PRISM

    The uncompromising power of Base, packed into the lightning speed of Turbo.

    ZIT-Prism (Z-Image-Turbo Prism) is a custom fine-tuned and fully distilled derivative of the Z-Image-Base model. If you love the deep prompt adherence, complex textures, and high-fidelity details of the Base model but hate waiting for 30–50 steps, this is the model for you.

    By carefully fine-tuning and distilling the Base architecture, ZIT-Prism achieves stunning, production-ready generations in just 8 steps.

    ✨ Why use ZIT-Prism?

    • Best of Both Worlds: Retains the rich compositional intelligence and aesthetics of Z-Image-Base while generating at the speed of Z-Image-Turbo.

    To get the absolute best results out of this distilled model, please use the following parameters:

    • Steps: 8

    • CFG Scale: 1.0

    • Sampler: Euler / Euler a

    Feel free to experiment!

    🧩 LoRA Compatibility (Important!)

    Because ZIT-Prism is fundamentally built on the Base architecture (despite its Turbo-like speed), only Z-Image-Base LoRAs are officially supported.

    • ✅ Z-Image-Base LoRAs: Work flawlessly.

    • ❓ Z-Image-Turbo LoRAs: You can certainly give them a try, but results may be unpredictable or deep-fried due to the conflicting distillation methods.

    💬 Prompting Advice

    ZIT-Prism thrives on both natural language prompting and tags as the model Z-Image was trained with both!

    As per this paper by Tongyi-MAI - creator of Z-Image: Paper page - Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer

    Any support you can give will very much be appreciated, as they will help me in my future endeavors as a creator, and also as a college student: https://ko-fi.com/ppoyaa

    Description

    Experimented for a bit, leading to a version 1.7 that I was pretty much satisfied with. Could update in the future.

    FAQ

    Comments (8)

    wrOngplanetMay 17, 2026
    CivitAI

    The text encoder here has the same SHA256 hash as another one I already had (named "zImageBase_textEncoderFP8.safetensors").

    SHA256 is:

    72450b19758172c5a7273cf7de729d1c17e7f434a104a00167624cba94f68f15

    Just FYI

    wrOngplanetMay 17, 2026

    Btw I don't know what's the "official" name of said text encoder, could be I found it under another name at some point. (Edit posts don't seem to work, so commenting here).

    aueki4g467May 18, 2026

    I can't remember when it started, but CIVITAI has been arbitrarily renaming downloaded files to whatever they decide. It's possible that the same file might have a different name if it was downloaded at a different time.

    amazingbeautyMay 17, 2026
    CivitAI

    FP16 unet pelase

    schschMay 18, 2026
    CivitAI

    Strange enough, this Diffusion Model in particular doesnt work with my current workflows using ZIT. It becames very blurry like made with 2 steps only in ksampler. I tried Euler, er_SDE, res_2m, with 8, 10, 20 steps, cfg1, cfg, cfg3, everyting. I even tried to convert to FP8 by using special nodes (it went down to 5.73gb but still has very blurry images). I am not using any custom LORA. What happened? Perhaps ist based on ZIB (Z-image Base) instead of ZIT being the culprit of not working in certain ComfyUI portable versions? I use ComfyUI 0.18.1, Pyhon 3.12.9

    My command line is: main.py --windows-standalone-build --lowvram --disable-auto-launch --fast fp16_accumulation --disable-smart-memory --disable-cuda-malloc --max-upload-size 1000

    Cryhell
    Author
    May 21, 2026

    I'm not too sure. I'm using a very basic workflow to generate. Basic Ksampler as well with just euler/simple and there's been no problem.

    obinna7713May 20, 2026
    CivitAI

    this model is very broken? all i get is blurry results especially the vae decoding phase.

    Cryhell
    Author
    May 21, 2026

    It works fine with the basic workflow I'm currently using. Try testing it with euler/simple?

    Checkpoint
    ZImageBase
    by Cryhell
    Download (Beta) View on CivitAI
    base model

    Details

    Downloads
    167
    Platform
    CivitAI
    Platform Status
    Available
    Created
    5/17/2026
    Updated
    5/23/2026
    Deleted
    -

    Files

    ae.safetensors

    Size:
    319.77 MB
    SHA256:
    afc8e28272cd15db3919bacdb6918ce9c1ed22e96cb12c4d5ed0fba823529e38

    Mirrors

    HuggingFace (602 mirrors)
    CivitAI (13 mirrors)
    TensorFiles (1 mirrors)
    ModelScope (7 mirrors)
    ModelScope CN (11 mirrors)
    ShakkerAI (1 mirrors)

    zitPrism_v17_txt.safetensors

    Size:
    5.25 GB
    SHA256:
    72450b19758172c5a7273cf7de729d1c17e7f434a104a00167624cba94f68f15

    Mirrors

    HuggingFace (41 mirrors)
    CivitAI (2 mirrors)
    ModelScope (2 mirrors)
    ModelScope CN (1 mirrors)

    zitPrism_v17.safetensors

    Size:
    11.46 GB
    SHA256:
    2f9db9446b00ef4c2942eb74528cba9e05e286065679037ca815e898840ea67b

    Mirrors

    CivitAI (1 mirrors)