Skip to main content
Control various parameter settings for image generation to help you create ideal image effects.

Image Generation Parameter List

Example Usage

Advanced Prompt Techniques

Multi-Prompts

Use double colons :: to separate concepts and control their relative importance:
The first example treats “Hot dog” as one concept, while the second separates “Hot” and “dog” as different concepts.

Prompt Weights

Add weights to multi-prompt sections using :: followed by a number:
This gives “Beach” twice the weight of “mountain forest” in the generation. Weight Examples:
  • ::2 - Double weight
  • ::0.5 - Half weight
  • ::-0.5 - Negative weight (reduces influence)

Negative Prompting Examples

Advanced Parameter Combinations

Ultra-detailed realistic photo:
Artistic and experimental:
Character consistency across images:
Anime style with motion:

Supported Aspect Ratios

Core aspect ratios and their typical applications:
  • 1:1 - Standard square, suitable for social media avatars, Instagram posts
  • 4:3 - Traditional rectangle, suitable for older monitors, tablets
  • 2:3 - Classic photography ratio, commonly used for frames, posters
  • 16:9 - Modern widescreen ratio, mainstream standard for HD video, smart TVs
  • 9:16 - Portrait ratio, optimized for mobile short videos, social media stories

Setting Aspect Ratios

You can quickly determine and set image aspect ratios through these methods:
  • Based on pixel dimensions: For 1920x1080 pixel images, input --ar 1920:1080, system automatically simplifies to --ar 16:9
  • Based on physical dimensions: For sizes like 8.5x11 inches, remove decimal points and input --ar 85:110
  • Custom ratios: Supports any integer ratios, such as --ar 3:2 or --ar 5:4

Important Notes

  • Default aspect ratio is 1:1, suitable for most basic scenarios
  • Only supports integer ratios, use 139:100 instead of 1.39:1
  • Different model versions may have varying support for aspect ratios
  • Extreme aspect ratios (like 1:10 or 10:1) are experimental features with potentially unstable results
  • During HD processing, some aspect ratios may be fine-tuned for optimization

Detailed Parameter Guidelines

Stylize Parameter Effects

Low Stylize (0-50): More literal interpretation of prompts
  • --s 0: Very literal, follows prompt closely
  • --s 25: Slight artistic interpretation
  • --s 50: Balanced between literal and artistic
Medium Stylize (100-250): Default range with good artistic balance
  • --s 100: Default value, good balance
  • --s 200: More artistic interpretation
  • --s 250: Strong artistic styling
High Stylize (300-1000): Heavy artistic interpretation
  • --s 500: Very artistic, may deviate from prompt
  • --s 750: Extremely artistic
  • --s 1000: Maximum artistic interpretation, minimal prompt adherence

Chaos Parameter Effects

Low Chaos (0-25): Consistent, predictable results
  • --c 0: Default, very consistent results
  • --c 15: Slight variation
  • --c 25: Moderate variation while staying on-theme
High Chaos (50-100): Unpredictable, varied results
  • --c 50: Significant variation
  • --c 75: High variation, experimental
  • --c 100: Maximum chaos, very unpredictable results

Quality Parameter Impact

  • --q 0.25: 4x faster, lower detail (draft quality)
  • --q 0.5: 2x faster, moderate detail
  • --q 1: Default quality and speed
  • --q 2: 2x slower, maximum detail (only v4+)

Weird Parameter Effects

Subtle Weird (0-250): Slight oddness
  • --w 0: Default, no weirdness
  • --w 100: Slightly unusual elements
  • --w 250: Noticeable strange elements
Strong Weird (500-1500): Significant strangeness
  • --w 750: Very unusual and surreal
  • --w 1500: Extremely weird and abstract
Maximum Weird (2000-3000): Extreme experimental results
  • --w 2500: Highly abstract and bizarre
  • --w 3000: Maximum weirdness, often unrecognizable

Seed Parameter Usage

Seeds are useful for:
  • Consistency: Same seed + prompt = similar results
  • Variations: Change other parameters while keeping seed
  • Experimentation: Test different prompts with same seed
  • Reproducibility: Share exact results with others
Important Notes:
  • Seeds only work with same model version
  • Different aspect ratios may produce different results even with same seed
  • Seeds from older model versions won’t work with newer versions

Parameter Combination Suggestions

  • High Quality Detailed Images: --quality 1 --stylize 250 --chaos 10
  • Creative Experimentation: --chaos 75 --stylize 50 --weird 500
  • Character Consistency: --cref [character image] --seed [fixed value] --chaos 0
  • Style Matching: --sref [style image] --stylize 100 --iw 1.5
  • Realistic Photography: --stylize 50 --quality 1 --raw
  • Abstract Art: --stylize 750 --weird 1000 --chaos 80
  • Anime/Manga Style: --niji 6 --stylize 200 --quality 1
  • Multiple Variations: --chaos 25 --stylize 150

Current Limitations

The API currently does not support the following parameters for image generation:
  • --repeat parameter
  • --bs, --batchsize parameter (only supported in video functions with values [1,2,4])
  • --personalize parameter
  • {} prompt variation syntax
  • --stealth/--public parameters
  • --video parameter (use video-specific endpoints instead)

Object Reference (—oref) - V7 Update

Object Reference allows you to reference an object from another image in your generation.

Usage

Add the --oref parameter at the end of your prompt, followed by your image URL. You can only use one image in object reference. Example:

Important Limitations (V7)

  • Incompatible with V6.1 features: Object reference is incompatible with features that still use V6.1 (such as inpainting or outpainting)
  • Double GPU consumption: Using object reference consumes 2x GPU time compared to regular V7 images
  • Editing incompatibility: Object reference images are currently incompatible with regional inpainting, pan, or outpaint. To edit these images, use the Edit interface and remove image references along with any --oref or --ow parameters
  • Speed mode incompatibility: Object reference is currently incompatible with speed modes or --q 4

Important Usage Requirements

  • When using enhance: The original job must be created in draft mode with parameters --v7 --draft
  • Draft mode: Ultra-fast image prototyping (10x faster than normal mode) with slightly reduced quality but consuming only half the GPU resources. Use --draft parameter to activate.

Supported Speed Modes

Speed modes control the GPU type used for task execution, determining task execution time. The API provides fast mode and turbo mode, with relax mode currently not supported.

Fast Mode (--fast)

Fast mode is the default GPU speed mode where API requests are prioritized using GPU resources for processing. Usually completes image generation within 1 minute.
  • Usage: Activated with --fast parameter
  • Default: Used when no speed mode is specified
  • Processing Time: Typically under 1 minute for standard images

Turbo Mode (--turbo)

Turbo mode uses a high-performance GPU pool with processing speeds up to 4x faster than fast mode. This mode consumes double the GPU time quota and only supports v5 and above versions.
  • Usage: Activated with --turbo parameter
  • Speed: Up to 4x faster than fast mode
  • Requirements: Only supports v5+ versions
  • Cost: Consumes double GPU time quota
  • Status: Experimental feature - availability and billing may change

Prompt Best Practices

Effective Prompt Structure

  1. Subject: What you want to see (e.g., “portrait of a woman”)
  2. Style/Medium: Art style or medium (e.g., “oil painting”, “digital art”)
  3. Details: Specific details (e.g., “blue eyes, flowing hair”)
  4. Environment: Setting or background (e.g., “in a garden”, “studio lighting”)
  5. Parameters: Add parameters at the end
Example:

Parameter Order Recommendations

Place parameters in this order for best results:
  1. Aspect ratio (--ar)
  2. Quality (--quality)
  3. Stylize (--stylize)
  4. Chaos (--chaos)
  5. Weird (--weird)
  6. Speed mode (--fast, --turbo)
  7. Model version (--v, --niji)
  8. Special modes (--raw, --tile, --draft)

Common Mistakes to Avoid

  • Too many conflicting parameters: Keep combinations logical
  • Extreme parameter values: Start with moderate values and adjust
  • Inconsistent aspect ratios: Choose ratios that match your intended use
  • Overuse of chaos: High chaos can make results unpredictable
  • Wrong model versions: Use appropriate version for your style needs