Z-Image Base Guide to Guidance, Negative Prompts and Seeds

Z-Image Base supports guidance and negative prompts for controlled image generation. Start with 28–50 steps and guidance of 3.0–5.0 using the official Tongyi-MAI/Z-Image checkpoint. Base model card

What are the key facts about Z-Image Base?

Decide which variable the experiment will test: prompt wording, guidance strength, exclusions, or seed. Keep the other settings fixed until you have compared the outputs.

What controls does Z-Image Base support?

The developer positions the undistilled release for varied styles, diverse compositions, negative-prompt control, and downstream development. Its model card also identifies it as a starting point for adaptation, including LoRA training. Treat these as documented design goals to test against your own brief. Base model card

For a useful trial, choose a scene where correctness is visible without a subjective style debate. A blue mug to the left of a red notebook provides concrete objects, colors, and a spatial relationship. Add a visual treatment only after checking those requirements.

Then consider diversity. Request the same scene across several seeds and inspect whether your approved requirements remain stable while the composition changes. Keep examples that fail as well as examples that pass. This suggested review method helps distinguish useful variation from an image that simply ignored part of the request.

The sibling Z-Image in ComfyUI guide covers component placement and template loading. Use it when you want to conduct the same kind of experiment through a visual graph.

What guidance and sampling limits affect Z-Image Base?

The Base card recommends 28–50 inference steps and guidance between 3.0 and 5.0. These are sampling recommendations, not a promise of a particular render time. Start within the published recipe before exploring settings outside it. Base model card

Diffusers explains that increasing guidance encourages closer alignment with a prompt, potentially at the expense of image quality. This creates a tradeoff to inspect rather than a setting to maximize automatically. A stronger instruction signal does not remove the need to check the finished image. Pipeline reference

The ZImagePipeline implementation enables guidance above zero. Use the Base card's recommended 3.0–5.0 range when testing negative prompts. Pipeline implementation Base card

The repository lists Z-Image Base for generation and Z-Image-Edit for editing. Keep the exact checkpoint identifier with your results. Project repository

Record memory, precision, resolution, and loaded components for your trial. Confirm that one generation completes before scheduling a larger experiment.

How do you test Z-Image Base guidance, prompts, and seeds?

Prepare a Python environment with PyTorch appropriate to your CUDA system and the dependencies described in the model card. The developer recommends installing Diffusers from its source repository. Once that environment is ready, this compact example follows the Base card's pipeline and uses settings within its recommended range. Base model card

import torch
from diffusers import ZImagePipeline

pipe = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image", torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=False,
).to("cuda")
image = pipe(
    prompt="A blue mug left of a red notebook on a pale desk",
    negative_prompt="printed lettering",
    width=1024, height=1024,
    num_inference_steps=28, guidance_scale=4.0,
    cfg_normalization=False,
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]
image.save("z-image-base.png")

The prompt, negative prompt, seed, and output name are example choices for this exercise. They are not recommended universal values. Save the result and write a short inspection note: which requirements passed, which failed, and whether the image is useful for the intended placement.

Next, perform a guidance comparison. Keep the prompt, seed, dimensions, and step count fixed. Select another guidance value within the published range and compare the outputs side by side. Note both instruction following and unwanted changes in appearance; do not evaluate only whichever criterion favors one image.

Now return to the original guidance setting and test the negative prompt separately. Compare an empty negative prompt with one specific exclusion. In this exercise, look for unwanted printed lettering. Avoid inserting a long generic exclusion list before you know which visible problem you are trying to solve.

Finally, vary seeds while keeping the chosen prompt and settings. Review the set as a collection. Does each image contain the right objects? Is their relationship understandable? Which variations would fit the intended layout? A seed is useful experiment metadata, but the selected image remains the actual deliverable.

Keep a result sheet with the prompt revision, seed, settings, passed requirements, and manual corrections. Explain why each image was accepted.

If your project later needs custom training, preserve this unmodified baseline. Evaluate a proposed adaptation against the same original brief and a separate set of new scenes. The LoRA guide provides the broader workflow context for loading and troubleshooting compatible adapters.

How does Z-Image Base compare with Turbo?

Z-Image-Turbo is the family's distilled alternative, optimized for a short generation schedule. Base is the alternative when you want the developer's undistilled model and its guidance controls. The published cards describe these different objectives; an actual creative comparison should use each model's own documented configuration. [Turbo card][turbo] Base card

Do not force identical sampling settings merely to make a comparison appear fair. Instead, hold the creative brief and output-review criteria constant. Record each complete recipe so that another person can understand the practical tradeoff you observed.

What should you know before configuring Z-Image Base?

What is the official Z-Image Base download name?

Download Z-Image Base from Tongyi-MAI/Z-Image on Hugging Face. Use that identifier with ZImagePipeline to load the undistilled model. Base card

Does Z-Image Base support negative prompts?

Z-Image Base supports negative prompts with guidance enabled. The pipeline activates guidance above zero; start with the recommended 3.0–5.0 range. Implementation Base card

Which Z-Image Base settings should I start with?

The Z-Image Base model card recommends 28–50 inference steps and guidance of 3.0–5.0. Keep the prompt and seed fixed while comparing guidance values within that range. Base card

Can I train a LoRA on Z-Image Base?

Z-Image Base is documented as a foundation for LoRA training and other adaptations. Choose a training implementation that supports this exact checkpoint. Base card

Sources

• Base model card
• Project repository
• ComfyUI guide
• Pipeline reference
• Pipeline implementation
• [Turbo card][turbo]

[turbo]: https://huggingface.co/Tongyi-MAI/Z-Image-Turbo<!-- pz-related-guides -->

Related guides on PromptZone

• Best SDXL Models in 2026
• ComfyUI 2026: The Complete Guide
• AI Model Releases Timeline