Best Schedulers and Samplers for Z-Image Base: res_2s and bong_tangent

Choosing the right scheduler and sampler can dramatically affect Z-Image Base's output quality and generation speed. While default settings work, optimized combinations like res_2s scheduler with specific samplers can produce noticeably better results. This guide explores the best scheduler and sampler configurations based on community testing and practical experience.

Understanding how schedulers and samplers work together helps you fine-tune generation for your specific needs.

Understanding Schedulers and Samplers

Before exploring specific recommendations, let's clarify what these components do.

What is a Scheduler?

The scheduler controls how noise is added and removed during the diffusion process:

Noise Schedule: Defines how noise levels change across generation steps Step Progression: Controls the pace of denoising Quality Impact: Affects detail resolution and clarity

Different schedulers create different "paths" through the denoising process, leading to varied results.

What is a Sampler?

The sampler determines how the model navigates from noise to image:

Navigation Method: How predictions are made at each step Convergence: How quickly the image stabilizes Characteristics: Deterministic vs stochastic behavior

Samplers affect both quality and the character of outputs.

How They Interact

Scheduler + Sampler combinations create unique generation characteristics. Some pairings work better than others, and optimal combinations vary by model architecture.

Top Scheduler: res_2s

The res_2s (resolution 2-step) scheduler has emerged as a favorite for Z-Image Base.

What Makes res_2s Special

Refined Noise Schedule: res_2s uses a two-stage resolution approach that better preserves detail during denoising. The schedule is tuned for transformer-based diffusion models.

Quality Benefits:

• Sharper fine details
• Better texture preservation
• Reduced artifacts
• Improved color accuracy

Efficiency:

• Achieves quality at lower step counts
• Faster convergence than some alternatives
• Good quality even at 20 steps

res_2s Settings

In ComfyUI:

Scheduler: res_2s
Steps: 25-30 (optimal range)
CFG: 6-7 (Z-Image Base sweet spot)

Why res_2s Works Well with Z-Image Base

The S3-DiT architecture responds well to res_2s's refined noise schedule. The model's sliding-window attention benefits from the scheduler's approach to resolution preservation.

res_2s maintains detail better through the denoising process

Alternative Scheduler: bong_tangent

The bong_tangent scheduler offers another excellent option for Z-Image Base.

What is bong_tangent?

A community-developed scheduler that uses a tangent-based noise schedule:

Smooth Transitions: Gradual noise reduction with minimal abrupt changes Creative Range: Tends to produce more varied outputs Good Balance: Quality without over-sharpening

bong_tangent Characteristics

Pros:

• Very smooth gradients
• Good for artistic styles
• Less prone to over-sharpening
• Interesting variations

Cons:

• May need more steps for maximum detail
• Slightly softer than res_2s
• Less common in workflows

bong_tangent Settings

Scheduler: bong_tangent
Steps: 30-35 (benefits from more steps)
CFG: 5-6 (slightly lower than res_2s)

When to Choose bong_tangent

• Artistic and painterly styles
• When softer transitions are preferred
• Creative exploration
• Avoiding over-sharpened outputs

Best Samplers for Z-Image Base

Pairing the right sampler with your scheduler matters.

Top Sampler: euler_ancestral

Why it works:

• Good balance of quality and speed
• Adds controlled randomness for variety
• Works excellently with res_2s
• Reliable across prompt types

Settings:

Sampler: euler_ancestral
eta: 1.0 (default)

Alternative: dpmpp_2m_sde

Why it works:

• Excellent detail preservation
• Smoother than euler in some cases
• Good convergence characteristics
• Works well with bong_tangent

Settings:

Sampler: dpmpp_2m_sde

Other Quality Options

dpmpp_3m_sde:

• Three-step method for refined results
• Slightly slower but higher quality potential
• Good for final renders

euler:

• Simpler, deterministic variant
• Consistent results across seeds
• Faster than ancestral variants

Sampler Comparison

Optimal Combinations

Tested combinations that work well together.

Combo 1: Quality Standard (Recommended)

Scheduler: res_2s
Sampler: euler_ancestral
Steps: 28
CFG: 7

Best for: General purpose, reliable quality

Combo 2: Maximum Quality

Scheduler: res_2s
Sampler: dpmpp_3m_sde
Steps: 35
CFG: 6.5

Best for: Final renders, portfolio pieces

Combo 3: Creative Exploration

Scheduler: bong_tangent
Sampler: euler_ancestral
Steps: 32
CFG: 5.5

Best for: Artistic styles, variation generation

Combo 4: Speed Priority

Scheduler: res_2s
Sampler: euler
Steps: 20
CFG: 7

Best for: Rapid iteration, previews

Different combinations produce different characteristics

CFG Scale Optimization

CFG (Classifier-Free Guidance) scale significantly affects output.

Z-Image Base CFG Sweet Spot

Unlike SDXL which often uses CFG 7-9, Z-Image Base works best at slightly lower values:

CFG by Use Case

Portraits: 6-7 Landscapes: 5-6 Abstract: 4-5 Product shots: 7 Text rendering: 7-8

Step Count Guidelines

How many steps you need depends on scheduler and desired quality.

With res_2s

With bong_tangent

ComfyUI Implementation

Setting up optimal configurations in ComfyUI.

KSampler Node Settings

Model: z-image-base
Seed: [your seed]
Steps: 28
CFG: 7
Sampler: euler_ancestral
Scheduler: res_2s
Denoise: 1.0

Advanced KSampler (Recommended)

Using the Advanced KSampler provides more control:

add_noise: enable
noise_seed: [seed]
steps: 28
cfg: 7
sampler_name: euler_ancestral
scheduler: res_2s
start_at_step: 0
end_at_step: 28
return_with_leftover_noise: disable

Custom Scheduler Nodes

For res_2s and bong_tangent, you may need custom nodes:

• Check ComfyUI Manager for scheduler packs
• Community nodes often include newer schedulers
• Some may require manual installation

Troubleshooting

Common issues with scheduler/sampler settings.

Over-Sharpened Output

Cause: CFG too high or wrong scheduler Solution: Lower CFG to 6, try bong_tangent

Muddy/Soft Results

Cause: CFG too low or insufficient steps Solution: Increase CFG to 7, add steps

Inconsistent Quality

Cause: Sampler variance or wrong combination Solution: Use euler instead of ancestral for consistency

Artifacts and Noise

Cause: Bad scheduler/sampler pairing or CFG issues Solution: Try res_2s + euler_ancestral at CFG 6-7

Frequently Asked Questions

What's the single best scheduler for Z-Image Base?

res_2s for most users. bong_tangent for artistic work.

Do I need custom nodes for these schedulers?

res_2s may be built-in depending on ComfyUI version. Check Manager for updates.

Can I use SDXL schedulers with Z-Image Base?

Some work, but model-optimized schedulers like res_2s perform better.

Why is my CFG different from SDXL recommendations?

Z-Image Base's architecture responds differently. Lower CFG values work better.

How many steps is too many?

Beyond 40-50 steps, quality improvements are minimal. 28-35 is the sweet spot.

Does sampler choice affect generation speed?

Slightly. euler is fastest, dpmpp_3m_sde slowest. Difference is small.

Should I change settings for different content types?

Minor adjustments help. Lower CFG for creative, higher for precision work.

What about negative prompts with these settings?

Negative prompts work normally. Settings don't change negative prompt behavior.

Are these settings compatible with LoRAs?

Yes, these scheduler/sampler settings work well with Z-Image Base LoRAs.

How do I install res_2s if it's not available?

Check ComfyUI Manager for scheduler node packs, or search GitHub for community nodes.

Optimizing scheduler and sampler settings elevates Z-Image Base from good to excellent. The res_2s and bong_tangent schedulers, paired with appropriate samplers and CFG settings, access the model's full potential.

For users who want optimal settings without manual configuration, Apatero offers pre-tuned Z-Image Base generation alongside 50+ other models, with LoRA training available on Pro plans.