How to Configure Negative Prompts, Guidance Scale, and Sampling Steps in Dezgo

Generating professional AI imagery requires more than writing an expressive prompt. When working with Dezgo—a versatile generation platform and REST API built around Stable Diffusion, SDXL, and Flux models—creators frequently run into common rendering issues: distorted hands, oversaturated colors, muddy textures, or AI outputs that wander far from instructions.

To tame these inconsistencies, Dezgo exposes three primary generation controls under its More options panel and in its API payloads: Negative Prompts, Guidance Scale (CFG), and Sampling Steps. Mastering how these three settings interact allows creators to systematically diagnose image flaws, sharpen prompt adherence, and optimize both generation speed and credit usage.

Where to Find the Core Parameters in Dezgo

Whether you are using Dezgo's free browser interface, pay-as-you-go Power Mode, or the REST API (POST https://api.dezgo.com/text2image), the controls are organized in a unified way:

1. Web Interface: On the generation page (such as Text-to-Image SD1 or Text-to-Image XL), locate and expand the More options accordion beneath the main prompt box. Here you will find the Negative Prompt input field alongside numerical sliders for Guidance and Steps.
2. API Payloads: When calling Dezgo endpoints programmatically, provide negative_prompt (string, up to 1,000 characters), guidance (float, range -20 to 20), and steps (integer, range 10 to 150).

Understanding what each parameter governs will prevent blind trial-and-error.

1. Negative Prompts: Eliminating Flaws and Unwanted Aesthetics

A positive prompt describes the concepts, lighting, and subjects the AI should construct in the latent space. A negative prompt tells the model which concepts, visual noise, and aesthetic signatures to steer away from during denoising.

The Baseline Starting Point

For standard Stable Diffusion 1.5 checkpoints (such as Absolute Reality or Realistic Vision), Dezgo provides a comprehensive default negative prompt:

ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, blurry, bad anatomy, blurred, watermark, grainy, signature, cut off, draft

This default acts as a safety net against common structural artifacts in earlier diffusion models. However, for specialized digital art or photorealistic renders, creators should customize the negative prompt to remove specific style interference (such as cartoon, 3d render, anime when seeking documentary realism, or photorealistic, oversaturated when designing flat vector art).

Dezgo's Prompt Weighting Syntax

When an unwanted element persists—such as unintended background bystanders or recurring extra fingers—Dezgo supports explicit prompt weighting syntax within both positive and negative fields:

• Increase Influence (+): Append one or more plus signs to boost suppression. For example, watermark+ or (extra fingers)++.
• Decrease Influence (-): Append minus signs if a negative term is overly aggressive and stripping desirable texture, such as grain-.
• Grouping: Enclose multi-word terms in parentheses before applying weight: (poorly drawn hands)+.
• Numeric Multipliers: Explicit numeric weights between 0 and 2 (where 1 is default) are supported using either colon or trailing format: (bad anatomy:1.2) or (bad anatomy)1.2.

Practical Rule of Thumb

Do not paste hundreds of generic negative buzzwords into every generation. Excessive negative tokens can dilute model attention and introduce unexpected flat colors or washed-out backgrounds. Keep negative prompts targeted to the visible flaws of the current subject.

2. Guidance Scale (CFG): Balancng Prompt Strictness and Image Coherence

The Guidance Scale (often known as Classifier-Free Guidance or CFG) determines how strictly the diffusion process is forced to conform to your text prompt relative to unconditioned image sampling.

In Dezgo, the guidance parameter accepts a technical range of -20 to 20, with a standard default set to 7.

Guidance Operating Zones

Architecture Caveats

Guidance requirements vary by model architecture available on Dezgo:

• Standard SD 1.5 & SDXL: Perform best between 6 and 9.
• SDXL Lightning Models (e.g., dreamshaperxl_lightning_1024px, juggernautxl_9_lightning_1024px): Lightning distillation checkpoints require significantly lower guidance values (typically 1.5 to 2.5) to avoid blown-out contrast.
• Flux Models (e.g., flux_1_schnell): Modern rectified flow models excel at natural prompt comprehension without aggressive guidance boosting.

3. Sampling Steps: Tuning Detail Depth Against Time and Cost

Diffusion models construct imagery through iterative denoising steps, transitioning an initial field of random gaussian noise into coherent visual geometry. The Sampling Steps parameter defines how many iterations this process undergoes.

Dezgo allows steps between 10 and 150. While more steps generally offer the sampler additional opportunities to resolve intricate textures, the benefits follow an aggressive law of diminishing returns.

Step Thresholds and Diminishing Returns

• 10 – 15 Steps (Rapid Drafts): Produces fast compositional mockups. The overall layout and color palette appear, but fine details (pupils, distant foliage, intricate jewelry) may remain smudged or blurry.
• 20 – 30 Steps (The Production Sweet Spot): High-efficiency standard for mainstream samplers like Euler or DPM++. The visual structure reaches mathematical convergence, and fine skin pores, fabrics, and lighting gradients lock into place.
• 35 – 50 Steps (Intricate Detailing): Beneficial for high-density visual subjects, such as detailed filigree, cyberpunk cityscapes, dense foliage, or macro photography.
• 60+ Steps (Diminishing Returns): Visual changes past 50 steps are virtually imperceptible in standard diffusion setups, but generation time increases linearly.

Account Balance and Free vs. Power Mode

In Dezgo, generation resources are managed differently depending on your tier:

• Free Mode: Has preset limits on maximum steps and resolutions to conserve shared server capacity.
• Power Mode & REST API: Functions on a prepaid pay-as-you-go credit balance. Because computing cost scales directly with step count and image dimensions, running 80 or 100 steps on an image where 25 steps would suffice depletes prepaid credits 3 to 4 times faster without discernible quality improvement.

Recommended Creator Workflow: The 4-Step Refinement Loop

Rather than adjusting all three levers simultaneously, follow this structured iteration workflow:

1. Phase 1: Exploration Draft
   • Keep Guidance at 7.0.
   • Set Steps to 20.
   • Use the default negative prompt.
   • Run small batches until you find a composition and Seed number you like.
2. Phase 2: Fix Structural Defects via Negative Weights
   • Lock the chosen seed.
   • Inspect the image for specific flaws: if hands are warped, add (poorly drawn hands)1.2 to the negative prompt; if text artifacts appear, add watermark+, text+.
3. Phase 3: Calibrate Prompt Adherence
   • If the AI is ignoring key descriptions (such as clothing color or specific background props), increment guidance from 7 to 8.5.
   • If skin textures look leathery or colors look radioactive, reduce guidance toward 6.0.
4. Phase 4: High-Detail Finalization
   • Once composition and fidelity are verified, increase Steps to 30–35 for a crisp final render.
   • Switch to Dezgo's upscale tool if high-resolution print or display output is needed.

By treating Negative Prompts as your scalpel, Guidance Scale as your adherence dial, and Sampling Steps as your convergence budget, you can reliably extract predictable, high-caliber artwork from Dezgo on every run.

References

• Dezgo - Free Online AI Image/Video Generator, Apps & Tools
• AI Image Generator API - Flux/Stable Diffusion - Dezgo
• Prompt weighting - AI Image Generator API - Flux/Stable Diffusion - Dezgo
• Models - AI Image Generator API - Flux/Stable Diffusion - Dezgo
• Overview - AI Image Generator API - Flux/Stable Diffusion - Dezgo