Use the Z-Image Turbo diffusion model, its Qwen text encoder and compatible VAE. Start from the basic workflow, test a clear natural-language prompt, then add editing or LoRA features only when needed.
Watch on YouTube · Video published 2025-12-02 · Companion reviewed 2026-09-23
Install the matching components
This lesson covers Z-Image Turbo, not every model with “Z-Image” in its name. Use the current official template and its linked model revisions. The split-file layout is:
- models/diffusion_models/z_image_turbo_bf16.safetensors
- models/text_encoders/qwen_3_4b.safetensors
- models/vae/ae.safetensors
Correction to the December 2025 video: the spoken instruction to put the split diffusion model in checkpoints is incorrect for this layout. Use models/diffusion_models. Model weights are separate downloads and are not included in the workflow pack.
Make the basic workflow work first
- Open the basic text-to-image graph. Choose the diffusion model, Qwen encoder and VAE in their matching loaders.
- Set a fixed seed and batch size 1. The video’s baseline uses 1024 × 1024, 8 steps, CFG 1, Euler sampling, simple scheduling and denoise 1. Treat this as that demonstrated graph’s starting point.
- Write a coherent scene description. Include the subject, its placement, the environment and the lighting. Run once before adding a LoRA or an upscale stage.
- Save the working baseline. Duplicate it for edits so you can return to a known configuration.
A natural-language starting prompt
Suggested example — not a reproduced video result
A matte green ceramic mug sits on a dark wooden desk beside a closed notebook. Soft morning light enters from a window on the left. The camera is close to the mug at tabletop height, with the notebook gently out of focus. The scene feels quiet and realistic.Describe relationships directly: which object is in front, where light comes from, and what should remain sharp. Natural language is a useful starting format for this model. It does not mean commas are forbidden, or that longer prompts always win. Compare a concise and a detailed version with the same seed.
Choose the workflow for the job
| Goal | Starting route | What to inspect |
|---|---|---|
| Make a new image | Basic text-to-image | Composition and prompt adherence |
| Rework an existing picture | Image-to-image | Denoise versus preservation of the source |
| Change a selected region | Inpainting | Mask edges and whether the graph composites untouched pixels |
| Add trained styling or a subject | Compatible Z-Image LoRA | Model compatibility, trigger words and strength |
| Finish an image | Upscale or sharpen branch | Texture, halos and unwanted invented detail |
Common mistakes
- A negative-prompt box is not proof that it affects the result. The CFG 1 Turbo baseline does not behave like a conventional SD1.5 negative-prompt setup.
- Do not load SDXL or FLUX LoRAs into Z-Image just because they use the same file extension. Match the model architecture.
- GGUF options can reduce memory requirements, but speed and image quality depend on quantization, node implementation and your hardware.
- For inpainting, test a small change first. Higher denoise can make a stronger change but may lose the original detail. Inspect the final composite rather than assuming all unmasked pixels are preserved.
Downloads and next step
The basic Z-Image workflow is free. The advanced workflow post is member content; Patreon shows the current access requirements. After you can repeat the baseline, use the prompt reference to compare image models without carrying incompatible settings across them.