CivArchive
    CielBleu ZIT - v2
    NSFW
    Preview 134820370
    Preview 134820738
    Preview 134820374
    Preview 134820599
    Preview 134820451
    Preview 134820447
    Preview 134820445
    Preview 134820885
    Preview 134820476
    Preview 134820454
    Preview 134820583
    Preview 134820463
    Preview 134820587
    Preview 134820481
    Preview 134820591
    Preview 134820469
    Preview 134820646
    Preview 134820464
    Preview 134820384
    Preview 134820649

    📝 Description

    CielBleu ZIT is an experimental fine-tuned merge, based on the original Z-Image Turbo and other ZIT models.

    It's designed for high-fidelity photorealism. It delivers photorealistic portraits with sophisticated lighting and organic texture rendering, maintaining full creative flexibility and prompt fidelity.

    Sampling Method : Euler
    Schedule Type : Simple
    Steps : 16
    CFG : 1
    Prompt : Long detailed natural language description, between 100-350 tokens.
    Resolution : Any width & height between 1024 and 2048, as long as they are divisible by 16.

    ⚠️ NSFW Note

    This model is capable of producing both SFW and NSFW content, but it has all the same flaws as the original Z-Image Turbo when it comes to NSFW content. I suggest either using a specialized NSFW Lora or performing SDXL inpainting.

    ℹ️ Required Additional Files

    Like the original Z-Image Turbo, CielBleu ZIT needs :
    Text Encoder : qwen3_4b.safetensors
    VAE : ae.safetensors

    📈 Versions

    v2 :

    - Slightly better NSFW capabilities
    - Improved anatomy consistency
    - Better prompts understanding / lighting logic
    - Faces are more natural
    - Available in both 16-bit (BF16 & FP16) and 8-bit (FP8)

    v1 :

    - Initial release
    - Only 8-bit (FP8)

    Description

    v2 :

    - Slightly better NSFW capabilities

    - Improved anatomy consistency

    - Better prompts understanding / lighting logic

    - Faces are more natural

    - Available in both 16-bit (BF16 & FP16) and 8-bit (FP8)

    FAQ

    Comments (19)

    dupe_900Jun 25, 2026
    CivitAI

    I wonder if he will be able to do Img2img. Preserve identities in the Img2img? Another question, if I have 14.7 Vram, can I run it? Or do I need 20 Gb of Vram?

    cielbleu9991
    Author
    Jun 25, 2026

    CielBleu ZIT does work in img2img, but a high denoise will change your character identity.

    If you really want to preserve identities, you should look into character LoRA, or inpaint only the body.

    Yes, you can run it with your 14.7 Gb of VRAM.

    Melodic_Possible_582589Jun 25, 2026· 1 reaction
    CivitAI

    automatic thumbs up

    Melodic_Possible_582589Jun 26, 2026· 1 reaction
    CivitAI

    so far so good. some of my prompts that originally generated borderline realism crossed into realism. This checkpoint seem to have some good stuff like nano banana pro or flux data mixed in or something like that. I would rate the at over 90% realism but under 99% realism because we all know this is ai so it can never be 100% :). I say "over" 90% because there still exists in some prompts the strong outline isolation of the subject to the background while some are fine. I think some more tweaking and this could be the most realistic checkpoint out there. Your goal is almost reached.

    cielbleu9991
    Author
    Jun 27, 2026

    Thank you @Melodic_Possible_582589 for the in depth feedback !

    TTTTTT959595Jun 26, 2026· 2 reactions
    CivitAI

    V2 is dope

    cielbleu9991
    Author
    Jun 26, 2026

    Thank you really much for your feedback.

    hatt2Jul 1, 2026
    CivitAI

    Hi @cielbleu9991 , any chance we will see your wonderful work in Krea2? 🙏🏼

    cielbleu9991
    Author
    Jul 1, 2026

    Maybe one day, even if it's unlikely.

    TTTTTT959595Jul 2, 2026

    +1, looking forward to seeing you bring your great talent to Krea2

    cielbleu9991
    Author
    Jul 23, 2026· 1 reaction

    Thank you for the support @hatt2 & @TTTTTT959595.
    I'm working on my Krea 2 checkpoint. I hope to release it soon.

    hatt2Jul 23, 2026

    @cielbleu9991 awesome!! thank u!

    lobliquaJul 6, 2026
    CivitAI

    For some reason V2 seems less consistent based on my tests. I tested all three versions (fp8, fp16, and bf16). The lighting is a major strong point of this version. I usually need to convert the models to a format recognized by the diffusers library for my Python workflow, and perhaps that is slightly altering the final result. However, just to share some feedback: V1 remains the best version for me in terms of stability and output quality, it is simply sensational.

    cielbleu9991
    Author
    Jul 7, 2026· 1 reaction

    Thank you for the feedback.
    To be honest, CielBleu ZIT was only tested on Forge Neo and ComfyUI simple workflows.
    I'm sorry you're experiencing instability with v2 in your Python workflow.

    lobliquaJul 7, 2026· 1 reaction

    @cielbleu9991 No problem, your work is simply fantastic! Thank you very much, and keep it up!

    12596936Jul 8, 2026
    CivitAI

    Why not releasing a higher precision than FP8 type?

    cielbleu9991
    Author
    Jul 8, 2026

    FP16 and BF16 were not possible with CielBleu ZIT v1.
    That's why I released CielBleu ZIT v2 on June 25, 2026, an improved version, available in FP8, FP16 and BF16.

    StrawberryicecakeJul 13, 2026· 1 reaction
    CivitAI

    Thanks! This ZIT model does female genitals a lot better than top-downloaded ZiT models that advertise with NSFW.

    cielbleu9991
    Author
    Jul 14, 2026

    Thank you really much.

    Checkpoint
    ZImageTurbo
    by cielbleu9991
    Download (Beta) View on CivitAI
    cielbleu zit
    z image turbo
    z-image
    nsfw
    realistic
    photorealistic
    base model
    cielbleu

    Details

    Downloads
    1,548
    Platform
    CivitAI
    Platform Status
    Available
    Created
    6/25/2026
    Updated
    8/3/2026
    Deleted
    -

    Files

    cielbleuZIT_v2_txt.safetensors

    Size:
    7.49 GB
    SHA256:
    6c671498573ac2f7a5501502ccce8d2b08ea6ca2f661c458e708f36b36edfc5a

    Mirrors

    HuggingFace (251 mirrors)
    CivitAI (14 mirrors)
    ModelScope (2 mirrors)
    ModelScope CN (4 mirrors)

    ae.safetensors

    Size:
    319.77 MB
    SHA256:
    afc8e28272cd15db3919bacdb6918ce9c1ed22e96cb12c4d5ed0fba823529e38

    Mirrors

    HuggingFace (721 mirrors)
    CivitAI (24 mirrors)
    TensorFiles (1 mirrors)
    ModelScope (9 mirrors)
    ModelScope CN (11 mirrors)
    ShakkerAI (1 mirrors)

    cielbleuZIT_v2.safetensors

    Size:
    11.46 GB
    SHA256:
    2a418b6ddae0276b25a277f6eb435e353edc209b66123e9b60033ffc60bc5ed9

    Mirrors

    CivitAI (1 mirrors)

    cielbleuZIT_v2.safetensors

    Size:
    11.46 GB
    SHA256:
    0a4cc16df1ed113486ed634969cf56a15c5838fc61aec9212bb603ac200b64d3

    Mirrors

    CivitAI (1 mirrors)

    cielbleuZIT_v2.safetensors

    Size:
    5.73 GB
    SHA256:
    7448afdd4b44b9aff6d5aac51e099ca3988968c6f0a5c3dc8f44b9ccf5719018

    Mirrors

    CivitAI (1 mirrors)