Choose between the split workflow or the all-in-one checkpoint approach.

1. Split workflow: Download the Z-Image diffusion model and place it in ComfyUI/models/diffusion_models. Download the Qwen text encoder and place it in ComfyUI/models/text_encoders. You also need a VAE – if you already work with Flux, you may already have it.
2. All-in-one checkpoint: Download the bundled file and place it in ComfyUI/models/checkpoints. This includes the model, text encoder, and VAE.
3. Check workflow notes for any required custom nodes. Install them via ComfyUI Manager if needed.

Correct node configuration is critical for Z-Image to interpret prompts properly.

1. Load CLIP node: Select the Lumina 2 type so the Qwen text encoder is interpreted correctly.
2. Model shift: The workflow commonly uses model shift 3.
3. Sampling settings:
   • Steps: Do not go above 9 steps – higher steps can degrade image quality.
   • Sampler and scheduler: Different combinations affect sharpness and skin rendering. Test with a fixed seed.
   • CFG: Use CFG 1 when no negative prompt is used – it keeps generation fast. CFG 2 or 3 increases saturation and contrast but slows generation. Lower CFG can look desaturated.
4. First run: Expect the first generation to be slower. Subsequent runs typically take 6–7 seconds on a high-VRAM system.

Z-Image’s output quality depends heavily on resolution. Use these techniques for Full HD and 2K.

1. Full HD (1920x1080): Direct generation can appear smoother or more diffuse. Instead, generate at an intermediate size that preserves detail, then use a Crop node after VAE Decode to remove extra pixels.
2. Portrait orientation: Swap width and height in the generation settings.
3. 2K output: Direct generation can distort the image. Generate at a suitable smaller resolution, then upscale the decoded image to 2K and pass it through a KSampler again with denoise 0.3 to refine details while preserving similarity.
   • Higher denoise (e.g., 0.7) creates a more different result – sometimes a desaturated cinematic matte look.
4. General upscale workflow: Use an upscale model to double the image, then refine at ~0.3 denoise. This can make plastic-looking Flux images more realistic, but may introduce artifacts – adjust seed and denoise.

Qwen VL can turn a reference image into a detailed prompt, or expand a short text prompt into a rich description.

1. Image to prompt: Load an image into the Qwen VL node, select a formula preset, and increase max tokens if the generated description is cut short. Feed the resulting long prompt into Z-Image.
2. Expand short text prompt: Concatenate a prompt formula with your simple input, send that to the Qwen VL custom-formula field, and connect the output to the text encoder. This creates richer prompt variation.

ControlNet ensures consistent structure across generations. Z-Image uses the Fun ControlNet Union model.

1. Download the Z-Image Fun ControlNet Union model and place it in ComfyUI/models/model_patches.
2. Loader node: This is in ComfyUI core. If missing or errors occur, update ComfyUI.
3. Preprocessors:
   • Canny: Extracts edges – good for keeping an illustration’s composition.
   • Depth Anything: Preserves depth and scene layout – may require several attempts with different seeds.
   • DW Pose: Extracts a pose skeleton – usually holds body position, but not tested on very complex poses.
4. ControlNet strength: 1.0 is too rigid in the demonstration. Use 0.6 to 0.8 for a good balance between control and creativity.

Train a LoRA to maintain a specific character across different outfits, backgrounds, and resolutions.

1. Dataset: Use at least 10 images (the example uses 15 consistent images of the same character from different angles). One image came from Z-Image; additional views were created with Nano Banana.
2. Training settings on Fal:
   • Steps: 1,000 steps
   • Learning rate: Default
   • Trigger token: Unique (e.g., chara_whitney)
   • Training type: Content (for character identity)
3. Cost: Approximately $2.26; completion time 15–20 minutes.
4. Download and load: Download the finished LoRA, name it after the trigger token, place it in ComfyUI/models/loras, press R to refresh, and load it with a LoRA loader.
5. Prompt: Put the trigger token plus the character description in the prompt. Explicit hair and eye colors improve identity accuracy.
6. LoRA strength: For non-photographic looks, reduce strength. 0.8 produced a much better watercolor result than full strength.

Style LoRAs require a coherent dataset – simply adding many images is not enough.

1. Dataset consistency: The first attempt with 40 images containing too many different subjects and styles failed at both 1,000 and 2,000 steps. A second dataset of 15 images restricted to game items worked well.
2. Training settings:
   • Steps: 1,500 steps
   • Learning rate: Default
   • Trigger token: Unique
   • Training type: Style
3. Testing: With the LoRA enabled, a fixed-seed 3D-looking object changed into the intended game-digital-painting style. Disabling the LoRA restored the original 3D appearance.