Z-Image Power Nodes v4 Complete Guide: 100+ Style Presets and Prompt Encoders
Why Power Nodes Is the "Style Powerhouse" of the Z-Image Ecosystem
Z-Image Turbo is the benchmark open-source image generation model for the speed-quality balance, but many users face a "happy problem": the model is so capable that reliably reproducing a specific artistic style becomes a struggle. Hand-writing style prompts is tedious, and consistency across different prompts is hard to maintain.
Z-Image Power Nodes (latest official release v2.0, referred to by the community as the v4/G2 generation) was built exactly to solve this. This open-source ComfyUI node suite by Martin Rizzo was distilled from his famous Amazing Z-Image Workflow, and has grown to 330+ stars and 450+ commits on GitHub with a "Very Positive" rating (70+ reviews) on CivitAI.
Its core value boils down to three points:
- 100+ preset style library: pick a style, and the node auto-generates the matching prompting — no manual tweaking
- Dedicated Z-Sampler Turbo sampler: composition consistency from just 3 steps; 7+ steps needs no post-processing
- Style consistency: keeps your subject and composition intact while shifting only the artistic direction — ideal for batch generation
Core Nodes Explained
⚡ Style & Prompt Encoder
This is the heart of the suite. It ships with 100+ predefined styles (covering illustration, photography, anime, photorealism and more), browsable through a searchable gallery with thumbnail previews. Once you pick a style, the node automatically rewrites your prompt, merges it with the style description, and encodes both using a CLIP text encoder.
Use case: You want to keep "an orange cat wearing a hat" as the subject and generate it in watercolor, oil painting, and cyberpunk styles — just switch the style parameter, and subject and composition stay unchanged.
⚡ Z-Sampler Turbo ^G2 (Second Generation)
The second-generation Z-Sampler Turbo rewrites the internal sampling logic from scratch. It keeps the first generation's 3-stage sigma-set design while adding new functionality and, more importantly, better parameterization of the internal sampling process. Two variants are offered:
| Variant | Highlights | Best For |
|---|---|---|
| Simple | Clean interface, all core features included | Most users (recommended) |
| Extended | More control parameters exposed | Advanced users, fine-grained control |
Low-step consistency: composition consistency holds from just 3 steps (minimal variation), results are acceptable at 5 steps, and 7+ steps deliver high quality with no post-processing needed. This lets you rapidly test prompts at low steps, then refine quality by raising the step count once the concept is locked.
⚡ Intensity and Intensity Bias
Key new parameters of the second-generation sampler:
- Intensity: tweaks the amplitude of the initial noise, affecting contrast and saturation
- Positive (>0.0): boosts contrast, sharpens edges, more defined and vibrant look
- Negative (<0.0): softer, "washed-out" look with less micro-detail
- Guidance: lower values suit photographic styles; higher values suit illustrations (prompt/style dependent)
- Intensity Bias: calibrates the noise bias; usually keep at 0.0
- Acts like a "brightness" adjustment, but the effect varies by prompt and style — it can even affect perceived focus
⚡ Turbo Creativity
Uses latent scrambling to increase variety across seeds, addressing Z-Image Turbo's limited seed variability. It only affects composition (posing, framing, object placement) — colors and style stay consistent. Note: it can cause hallucinations; in that case use the "refined" options, which add extra sampling steps for coherence (at the cost of generation time).
⚡ Other Utility Nodes
| Node | Purpose |
|---|---|
| Style String Injector | Injects the chosen style text directly into your prompt strings |
| My Top-10 Styles | One-click activation for your 10 most-used styles |
| VAE Encode (for Soft Inpainting) | Z-Image-optimized encoder that embeds the inpainting mask |
| Save Image | Saves images with optional CivitAI-compatible metadata |
| Empty Z-Image Latent Image | Creates correctly-sized latents; select aspect ratio, scale, orientation |
Installation Guide
Method 1: ComfyUI Manager (Recommended)
- Click "Manager" → "Custom Nodes Manager"
- Search "Z-Image Power Nodes"
- Click "Install", then restart ComfyUI
Method 2: Manual Installation
cd <ComfyUI directory>/custom_nodes
git clone https://github.com/martin-rizzo/ComfyUI-ZImagePowerNodes
Recommended Model Combinations
You need ae.safetensors (335 MB) in ComfyUI/models/vae/. Three model options:
| Model Type | Diffusion Model (→ diffusion_models/) | Text Encoder (→ text_encoders/) | Notes |
|---|---|---|---|
| GGUF (Q8/Q5) | z_image_turbo-Q5_K_S.gguf (5.19 GB) |
Qwen3-4B-Q8_0.gguf (4.28 GB) |
Best quality-per-GB, slower |
| FP8 Safetensors | z-image-turbo_fp8_scaled_e4m3fn_KJ.safetensors (6.16 GB) |
qwen3_4b_fp8_scaled.safetensors (4.41 GB) |
Native support, fast |
| BF16 Safetensors | z_image_turbo_bf16.safetensors (12.3 GB) |
qwen_3_4b.safetensors (8.04 GB) |
Default in official examples |
Author's guidance: if speed and native support weren't factors, GGUF offers the best quality for its size; for daily use, FP8 is recommended.
Hands-On: A Stylized Generation Workflow
Using "an orange cat wearing a hat" as the example, build a complete workflow:
- Empty Z-Image Latent Image: choose 1024×1024 or 16:9
- Style & Prompt Encoder: pick "Watercolor Illustration", enter your prompt
- Z-Sampler Turbo (Simple): Steps=7 (preview) or 12 (refine); adjust Intensity per style
- VAE Decode + Save Image: output and save with CivitAI metadata
Prompt: an orange cat wearing a hat, sitting on a windowsill, afternoon sunlight
Style: Watercolor Illustration
Steps: 7 (preview) / 12 (refine)
Intensity: 0.15
Batch Style Comparison Tip
Use the My Top-10 Styles node to add 10 candidate styles to favorites, then batch-generate with the same seed to quickly compare how the subject looks under each style. Pick the best one, then refine at higher steps.
FAQ
Q: Nodes don't show up after installation?
Cause: ComfyUI is outdated or not restarted.
Fix: Update ComfyUI to the latest version and fully restart; verify custom_nodes/ComfyUI-ZImagePowerNodes exists.
Q: How is Z-Sampler Turbo different from a normal KSampler?
Z-Sampler Turbo is purpose-built for Z-Image Turbo's distillation-optimized architecture: a 3-stage sigma set plus intensity control delivers far better consistency and quality at low steps than a generic KSampler. With a standard KSampler, use DPM++ 2M Karras with 8 steps.
Q: Turbo Creativity produces strange objects?
This is a known side effect of latent scrambling. Switch to "refined" mode for extra sampling steps, or disable the feature and change the seed manually.
Q: CivitAI can't read parameters from my saved image?
Confirm the CivitAI metadata option on the Save Image node is enabled; for complex workflows, append >>C to a node's title to force it into the metadata export.
Conclusion
Z-Image Power Nodes is the most essential custom node suite in the Z-Image / Z-Image Turbo ecosystem today. The 100+ preset style library solves the core pain point of style consistency, the second-generation Z-Sampler Turbo makes low-step high-quality generation the norm, and Intensity control plus Turbo Creativity give advanced users fine-grained tuning space.
Whether you're a casual user who wants fast results or a workflow tinkerer chasing batch style comparisons, this suite will noticeably boost your creative efficiency.
Next up: Z-Image Sampler and Scheduler Ultimate Guide — 14 Samplers × 10 Schedulers Benchmarked