CivArchive
    CyberRealistic Z-Image Turbo - v4.0
    NSFW
    Preview 125975265
    Preview 125978001
    Preview 125976638
    Preview 125976633
    Preview 125976755
    Preview 125976682
    Preview 125976686
    Preview 125978007
    Preview 125978165
    Preview 125976745
    Preview 125978010
    Preview 125978014
    Preview 125978160
    Preview 125978246
    Preview 125978068

    You can get this model through Civitai Early Access, or grab it along with many others by joining The Tinkerer on Whop. Membership gets you early releases, private tools and members-only pages. There are free pages too, no membership needed.

    👉 Join on Whop
    💬 Join the community for support, free tools and early news on Discord

    CyberRealistic Z-Image Turbo is a realism-focused finetune of Z-Image Turbo by Tongyi-MAI.

    The idea behind it is deliberately simple: keep what makes Z-Image Turbo good - speed, strong prompt understanding, good composition and extremely efficient few-step generation - while moving the default visual language further toward believable photography.

    CyberRealistic doesn't try to turn Z-Image Turbo into a completely different model. The original already has a very capable photographic foundation. The finetune mainly changes what the model considers a "normal" photograph: more natural skin, less synthetic rendering, more believable faces, stronger material texture, more grounded lighting and better anatomical consistency.

    Z-Image Turbo already knows how to make a good image. CyberRealistic mainly changes where it starts.

    What's different from base Z-Image Turbo

    • Stronger photographic look out of the box.

    • More natural skin texture with less waxy or overly polished rendering.

    • Improved faces, eyes, hair and small facial details.

    • Better anatomical consistency, especially hands, feet and complex poses.

    • Fewer duplicated limbs and extra hands in more difficult compositions.

    • More believable fabric, hair, skin, metal, glass and other material textures.

    • Stronger response to available light, practical lighting and real-world camera language.

    • Less dependence on stacks of words like masterpiece, 8k, ultra detailed and photorealistic.

    • Keeps the speed and general prompt behavior that make Z-Image Turbo useful.

    The focus is photography, but that doesn't mean the model is locked to photography. Illustration, cinematic stylization, fantasy, advertising, vintage photography and other looks are still available when you describe them.

    Prompting

    If you're coming from SDXL, Pony or Illustrious, the biggest change is simple:

    Describe the image instead of building a tag stack.

    Z-Image Turbo uses a Qwen3-based text encoder and responds very well to normal descriptive language. Short comma-separated clauses are completely fine, but every part of the prompt should ideally tell the model something visual.

    Instead of:

    woman, realistic, masterpiece, best quality, detailed skin, cinematic, 8k

    try:

    A woman standing beside an open apartment window on a warm summer evening, photographed with soft natural light falling across her face, loose dark hair, natural skin texture and an out-of-focus city street behind her.

    The second prompt gives the model an actual scene to construct.

    Put the subject first

    Start with what the image is about.

    A middle-aged mechanic leaning over the open engine bay of an old red pickup truck...

    works better than hiding the subject halfway through a long list of style instructions.

    You don't need to obsess over exact prompt order, but the main subject and composition should be clear early.

    Be specific

    Specific visual language usually does more than generic quality words.

    Instead of:

    beautiful lighting

    try:

    soft late-afternoon sunlight entering through a dusty workshop window

    Instead of:

    detailed clothing

    try:

    a faded blue denim jacket with worn seams and slightly frayed cuffs

    Instead of:

    cinematic portrait

    try:

    photographed from chest height with a 50mm lens, shallow depth of field and soft window light from camera left

    Describe the light

    Lighting is one of the easiest ways to change the realism and mood of the image.

    Useful examples:

    soft overcast daylight

    direct midday sunlight creating hard shadows

    a single warm tungsten lamp above the table

    cold fluorescent supermarket lighting

    late-afternoon sunlight entering through venetian blinds

    direct on-camera flash in a dark room

    You can still use words like cinematic, but describing where the light actually comes from gives the model much more information.

    Quality tags are not magic switches

    Words such as:

    masterpiece
best quality
8k
ultra detailed
absurdres
score_9

    can still influence the wording of the prompt, but Z-Image Turbo doesn't treat them like the traditional SDXL/Pony quality system.

    Use that prompt space to describe what you actually want to see.

    Camera language works well

    For photographic images, camera terminology can be useful when it describes a visible effect:

    35mm documentary photograph

    85mm portrait lens with shallow depth of field

    handheld photograph with slight motion blur

    direct flash snapshot

    wide-angle environmental portrait

    medium-format color photograph

    Don't feel forced to specify a camera and lens in every prompt. Sometimes simply saying casual phone photo gives you exactly the look you need.

    Prompt length

    There is no perfect prompt length, but these are useful practical ranges:

    • 10–30 words: exploration and seed hunting.

    • 30–80 words: good balance between control and freedom.

    • 80–150 words: complex scenes, precise lighting or detailed compositions.

    Long prompts aren't automatically better. Contradictory prompts are the bigger problem.

    If you ask for soft natural window light, hard direct flash, deep cinematic shadows and flat commercial studio lighting at the same time, the model still has to decide which instruction wins.

    Text inside images

    Z-Image Turbo is unusually capable at rendering text compared with older diffusion models.

    If exact text matters, put it in quotation marks:

    A small neon sign above the diner entrance reading "OPEN ALL NIGHT"

    Keep important text reasonably short. It's good, but it still isn't a replacement for a typography application.

    Z-Image Turbo is a distilled few-step model.

    Don't treat it like an SDXL checkpoint that needs 30–50 steps.

    A good starting point is:

    • Steps: 8–9

    • CFG / Guidance: effectively OFF

    • Resolution: start around 1 megapixel and increase if your hardware allows it

    • Negative prompt: normally unnecessary

    In the original Diffusers implementation, guidance is 0.0.

    In standard ComfyUI workflows, the equivalent no-CFG setup is generally CFG 1.0.

    More steps are not automatically better with Turbo. If something isn't working, changing the prompt, seed, sampler or composition usually makes more sense than simply increasing the step count.

    ComfyUI

    For ComfyUI, I recommend starting with the current Z-Image Turbo workflow/template rather than applying old SDXL settings.

    Z-Image Turbo has its own sampling behavior and is designed around very low step counts.

    Negative conditioning is normally zeroed out in the standard Turbo workflow because the model runs without traditional classifier-free guidance.

    Example prompts

    Natural-light portrait

    A woman in her early thirties sitting beside an open café window, loose brown hair falling across one side of her face, wearing a simple cream-colored sweater. Photographed from slightly below eye level with a 50mm lens, soft overcast daylight entering from the window, natural skin texture, muted colors and a busy street softly blurred in the background.

    Documentary photography

    An elderly fishmonger arranging silver mackerel on crushed ice at an indoor market early in the morning. Cold daylight enters through the open market doors and mixes with the warm bulbs above the counter. Wet concrete floor, weathered hands, faded rubber apron, handheld 35mm documentary photograph with subtle grain and natural color.

    Low-light snapshot

    A young woman standing alone beside a vending machine outside a convenience store at two in the morning, photographed with direct on-camera flash. Dark parking lot behind her, slightly messy hair, casual oversized jacket, realistic skin texture, hard flash shadows, muted colors and the imperfect look of a spontaneous late-night photograph.

    A few last things

    Short prompts are completely valid.

    One of the advantages of Turbo is that you can generate several directions quickly, choose the seed or composition you like, and then add more camera, lighting and material detail.

    That often works better than trying to write the perfect 150-word prompt before generating anything.

    Also keep in mind that Z-Image Turbo is distilled for speed. Part of that tradeoff is lower variation than a large non-distilled foundation model. If you keep seeing the same interpretation, change the wording more substantially rather than adding another five quality tags.

    CyberRealistic Z-Image Turbo is released for people who enjoy generating, experimenting, benchmarking and finding the edges of a model.

    Feedback is especially useful for difficult poses, multiple people, hands and feet, unusual lighting, text rendering and prompts where the model behaves differently from the original Z-Image Turbo.

    If you find something interesting — good or bad — let me know.

    Credits

    CyberRealistic Z-Image Turbo is based on Z-Image Turbo by Tongyi-MAI.

    Z-Image Turbo is released under the Apache 2.0 License. Please follow the applicable upstream license when using or redistributing derived models.

    Description

    Join The Tinkerer on Whop. Membership gets you early releases, private tools and a bunch of extra stuff.
    👉 Join on Whop

    This update brings further refinements across the board - better details and colours, improved consistency, and enhanced NSFW content.

    This release comes with +100 new sample images added to the already extensive V3.0 gallery, covering an even wider range of styles and scenarios.

    FAQ

    Comments (38)

    n00b_TGTMar 31, 2026· 2 reactions
    CivitAI

    "...After Early Access ends, downloads will be temporarily disabled." - Be nice to us users. We can still buy one copy and make it available for free.

    Cyberdelia
    Author
    Mar 31, 2026

    Just wait until tomorrow V3 will be available then.

    gurilagardnrMar 31, 2026· 14 reactions
    CivitAI

    I paid 5k buzz for v3, which, honestly, was sub-par. Now you want us to pay another 5k buzz for v4, in the hopes that it doesn't suck. yea...I get training isn't free, but, I think this business model sucks.

    PixelQueenMar 31, 2026· 7 reactions

    I think 5k is perfectly reasonable for these models; it sounds to me like your prompting is an issue if you think v3 was sub-par. I strongly disagree. I think v3 is an excellent model, and I am sure that v4 will be even better. If you do not wish to pay the very low sum of 5k for these models, then perhaps I suggest you wait until it becomes available.

    xFennxMar 31, 2026· 1 reaction
    CivitAI

    Does it work with character LoRAs yet?
    Do any Z-Image Turbo models work with character LoRAs other than the original?

    GetcontactApr 2, 2026

    I'm also looking for a custom checkpoint that's as stable as the bf16 original ZIT, but I still haven't found it. For all custom checkpoints, the weights of the models interfere with each other with the character lora. Maybe you've found something that's just as stable?

    Cyberdelia
    Author
    Apr 3, 2026

    @xFennx @Getcontact I have released https://civitai.com/models/2513307/cyberrealistic-z-image-turbo-catalyst (CyberRealistic Z-Image Turbo Catalyst) which work (better) with character LoRa's.

    GetcontactApr 3, 2026

    @Cyberdelia wow! I have a lot of random character lora and today i will test Catalyst model, thanks! If I find the optimal "sampler/scheduler" for my lora's and your Catalyst and get good images, I will definitely post them on the model's page.

    Cyberdelia
    Author
    Apr 3, 2026· 2 reactions

    @Getcontact There will be an alt version that is even better. Maybe this weekend.

    GetcontactApr 4, 2026

    @Cyberdelia I check this too, thanks!!

    ralphjxApr 18, 2026· 1 reaction

    @Cyberdelia Where is the alt version xD

    GetcontactApr 19, 2026· 1 reaction

    @ralphjx Catalyst version

    PixelQueenMar 31, 2026· 6 reactions
    CivitAI

    v3 was an outstanding model, and I am excited to try out v4. Thank you so much for all your hard work. I hope you continue well into the future; you inspire us all!

    larflarfApr 3, 2026
    CivitAI

    I'm not sure if I should ask this, but are these models Lora mix models, or are they models where you fine-tuned the Z-image?

    Cyberdelia
    Author
    Apr 3, 2026· 1 reaction

    At the moment, it is easy to convert custom training into LoRA blocks and merge them directly into the model. This works differently compared to other models like Z-Image Base, for example.

    nachyodaddyApr 3, 2026· 6 reactions
    CivitAI

    Required Additional Files

    Make sure you also have the following:
    16 GB+ VRAM: qwen3_4b.safetensors
    8–12 GB VRAM: qwen34bfp8_scaled.safetensors
    All VRAM sizes VAE: ae.safetensors


    New to ZIT and using forge neo, where exactly do i put these?

    Cyberdelia
    Author
    Apr 3, 2026· 1 reaction

    qwen3_4b in models/text_encoder
    ae in models/VAE

    Cyberdelia
    Author
    Apr 3, 2026

    If you need more help with Forge - just DM me. Always love to help and promote Forge Neo :)

    AFD_0Apr 3, 2026· 2 reactions

    I'm getting an error trying to download the Text Encoder from your HF folder (says "invalid username or password" and navigating down to the FluxTextEnc_VAE folder just returns a 404 error. Edit; Only getting this error for the FP8 text encoder, the full encoder link hosted at CivitAI works fine. (Nm, just drilled down on the full link at HF and found them both - wait, no, that says "mixed" not "scaled", so not sure if it's same thing or not).

    xJollyboyxApr 4, 2026

    i use the zimage vae and text encoder with it, works fine

    xJollyboyxApr 4, 2026

    @Cyberdelia with you on that!

    SaburoDioApr 20, 2026

    I am a bit confused by this as I have read it before. Same with some checkpoints.

    I run a RIG with a 4060 (12GB of VRAM) and 16GB of RAM.
    Never did I run into the slightest issues using the 16GB+ VRAM (qwen3_4b and BF16 checkpoint of Miraclein, BigLove etc).

    Maybe someone can explain?

    jesper123160Apr 3, 2026· 13 reactions
    CivitAI

    Let us know when it learns to generate penises.

    Cyberdelia
    Author
    Apr 3, 2026· 14 reactions

    Don't be a .... "dick". That feature is really "hard" ;)

    mvqsrtzApr 4, 2026

    Not being a jerk or promoting but you can try some loras or other nsfw-aimed models st. darkbeast/pornmaster. Honestly Z-Image (turbo) removed penis at the pre-train stage thus it is even inpossible to make it learn how dick looks like without thoroughtly retraining the whole model.

    pink0909Apr 6, 2026

    @mvqsrtz where did you find these insider inofs? do you know if the normal z-image (without turbo) still has penises? (i encountered this phenomenon on all SDXL turbo checkpoints (i tried) too the turbo seems to castrated models in general i dont know why)

    mvqsrtzApr 7, 2026· 2 reactions

    @pink0909 No. Z-image deformed all penises at the very beginning stage (likely in the trainning data), so all the z-image series (z-image-turbo or base) cannot generate good penis without deep modification. Probably a retrained z-image-base could fix it but I don't see that could happen very soon.

    ultraglidemailcom377Apr 7, 2026· 2 reactions

    Trying to be "cocky", eh? ;)

    girlswithafrosApr 23, 2026

    Moody porn merge does an "OK" wiener. Not great, but decent. This model was obviously not trained on wieners.

    Melodic_Possible_582589Apr 15, 2026
    CivitAI

    I will have to test this out this weekend. I am currently testing v3 and its amazing. Thanks for keeping the workflow in the images. I can just drag and drop the images to get the settings in Neo and do testing.

    playmusicApr 16, 2026
    CivitAI

    thanks for another great model. but in v4, i see more flow_dpo effect than needed. when i use it with dpo lora at around -0.7, your model clearly shines with er-sde / simple.

    jeremypierre076910Apr 21, 2026
    CivitAI

    Wow the complexity really shines over SDXL, Zimage is cooking.

    6vidit9Apr 23, 2026
    CivitAI

    INSANE REALISM, ABSOLUTE PEAKKK!!!

    JFTCApr 24, 2026· 9 reactions
    CivitAI

    So I was trying to make the switch to ZIT from SDXL but cyberrealistic models were just too good! Imagine my surprise when I checked and found that indeed my favorite creator also made ZIT models. God damn I use your models for everything now. Pony, SDXL, ZIT. Love you man.

    timstertimster3Apr 27, 2026
    CivitAI

    perfetly useful, tytyty

    PRO TIP: make low res image in SDXL then upscale and refine in ZiT.

    MGM77May 12, 2026
    CivitAI

    No matter what I try I cannot get the text encoder and VAE to work.... I am sure I stored them in the right location. Any advice?

    psylo1138287May 13, 2026
    CivitAI

    The link to qwen34bfp8_scaled.safetensors doesn't work...

    AddicteddMay 13, 2026· 4 reactions
    CivitAI

    Literally good at everything its nutty

    Checkpoint
    ZImageTurbo
    by Cyberdelia
    Download (Beta) View on CivitAI
    fantasy
    realistic
    artistic
    photorealistic
    base model

    Details

    Downloads
    6,978
    Platform
    CivitAI
    Platform Status
    Available
    Created
    3/31/2026
    Updated
    10/5/2026
    Deleted
    -

    Files

    cyberrealisticZImage_v40_txt.safetensors

    Size:
    7.49 GB
    SHA256:
    6c671498573ac2f7a5501502ccce8d2b08ea6ca2f661c458e708f36b36edfc5a

    Mirrors

    HuggingFace (269 mirrors)
    CivitAI (14 mirrors)
    ModelScope (3 mirrors)
    ModelScope CN (5 mirrors)

    ae.safetensors

    Size:
    319.77 MB
    SHA256:
    afc8e28272cd15db3919bacdb6918ce9c1ed22e96cb12c4d5ed0fba823529e38

    Mirrors

    HuggingFace (792 mirrors)
    CivitAI (24 mirrors)
    TensorFiles (1 mirrors)
    ModelScope (14 mirrors)
    ModelScope CN (12 mirrors)
    ShakkerAI (1 mirrors)

    cyberrealisticZImage_v40_pruned_fp8.safetensors

    Size:
    5.73 GB
    SHA256:
    1dea8e69a5bd51c2a772d81a7824e5875a0cc9e9f3a7c8be8889e220add59105

    Mirrors

    HuggingFace (4 mirrors)
    CivitAI (1 mirrors)

    cyberrealisticZImage_v40_full_bf16.safetensors

    Size:
    11.46 GB
    SHA256:
    052aaefdc7c973e51f23409f1ab722baefac9fd43ba4d4c4e054917ea7420b73

    Mirrors

    HuggingFace (2 mirrors)
    CivitAI (1 mirrors)