How to Configure Prompt Weights, Negative Prompts, and Dimensions on Lexica

Lexica is both a curated search engine for AI-generated imagery and a dedicated image generation suite powered by its proprietary Aperture model series. While standard text prompts provide the creative foundation for an image, mastering Lexica's generation controls—specifically canvas dimensions, negative prompts, and prompt weighting—allows creators to move beyond randomized results and achieve precise compositional framing, cleaner anatomical rendering, and consistent stylistic fidelity. Understanding how to configure these parameters in the Generate editor transforms prompt ideas into publication-ready visual assets.

Navigating Lexica's Generation Workspace

When accessing the Generate tab on Lexica, users are presented with a streamlined interface designed for rapid experimentation. Unlike complex open-source interfaces that require manual configuration of sampling steps, schedulers, and clip skip values, Lexica abstracts these low-level technicalities into core creative levers:

• Prompt Input Field: The primary text box where creators define the subject, artistic medium, lighting, environment, and visual atmosphere.
• Negative Prompt Field: An exclusion box dedicated to keywords and concepts that the diffusion process must actively suppress.
• Dimension Slider: A continuous or stepped selector that adjusts width and height simultaneously to set the output aspect ratio and pixel resolution.
• Guidance Scale Slider: The numerical control corresponding to Classifier-Free Guidance (CFG), determining how strictly the model follows prompt instructions versus hallucinating its own details.
• Model Selector: Allows toggling between Aperture model iterations (such as Aperture v2, v3, or v4), each calibrated with distinct photorealistic rendering styles and prompt sensitivities.

By methodically balancing these settings, creators eliminate the guesswork that often leads to distorted compositions, unwanted artifacts, or generic outputs.

Configuring Canvas Dimensions and Aspect Ratios

Canvas dimensions dictate not only the pixel count of the generated image but also its composition, subject framing, and spatial hierarchy. Lexica handles dimensions through an integrated aspect ratio control that updates the width and height parameters in tandem.

Impact on Composition and Framing

In diffusion-based models like Aperture, the aspect ratio establishes the geometric bounding box that guides subject placement:

• Portrait Orientations (e.g., 512x768 or 640x1024): Vertical framing encourages the model to generate full-body figures, standing portraits, and tall architectural features. If a prompt describes a character in a square frame, the model may crop the subject at the shoulders; switching to a portrait aspect ratio provides the necessary vertical canvas to render full garments, footwear, and grounded stances.
• Landscape Orientations (e.g., 768x512 or 1024x640): Horizontal framing suits panoramic vistas, cinematic stills, automotive designs, and multi-subject interactions. The wider aspect ratio distributes spatial depth across the horizontal axis, allowing for detailed background landscapes, ambient room sets, and lateral negative space.
• Square Orientations (e.g., 512x512 or 768x768): Square framing remains ideal for avatar portraits, close-up product renders, symmetrical architectural facades, and social media feed placements. It concentrates visual weight toward the center of the frame.

Avoiding Aspect Ratio Artifacts

Extreme aspect ratios can introduce unwanted repetition, such as duplicate heads, multiple torsos, or split horizons. This occurs when the model attempts to fill excessive canvas width or height with duplicate patterns from its training distribution. When generating wide or tall compositions:

1. Maintain standard aspect ratios (such as 2:3, 3:2, 16:9, or 9:16) supported directly within the Lexica slider interface.
2. If subject duplication occurs in wide images, incorporate compositional anchors in the positive prompt (such as "wide-angle panoramic view of a single solitary figure in the desert") to reinforce single-subject coherence across the horizontal span.

Eliminating Flaws with Negative Prompts

Negative prompts provide a mathematical steering mechanism within Classifier-Free Guidance. During each sampling step, the diffusion model calculates two vector paths: one moving toward the positive prompt conditioning and another moving toward the negative prompt conditioning. By subtracting the negative direction from the positive direction, Lexica suppresses specified features from materializing in the final render.

Core Negative Prompt Categories

Effective negative prompts address three primary categories of image degradation:

1. Anatomical and Structural Distortions: Diffusion models can struggle with complex joints and extremities. Supplying terms such as extra fingers, mutated hands, poorly drawn face, fused limbs, unnatural posture, cross-eyed signals the model to discard noisy latent steps that produce malformed anatomy.
2. Technical Artifacts and Compression Flaws: To ensure sharpness and tonal clarity, users frequently include blurry, low resolution, jpeg artifacts, pixelated, bad contrast, overexposed, underexposed.
3. Unwanted Text and Graphic Overlays: Because training datasets often include promotional material and stock photography, models may inject gibberish typography or watermark outlines. Using watermark, signature, text, logo, trademark, frame, border strips away unwanted commercial artifacts.

Best Practices for Negative Prompting

• Target Specific Defects: Add negative terms in response to observed issues in test generations rather than pasting hundreds of indiscriminate keywords. Overloading the negative prompt field can wash out image contrast, mute natural textures, and cause the subject matter to look artificially plasticky.
• Avoid Semantic Overkill: Negative prompting should not be used to remove essential elements that conflict with the positive prompt. For instance, putting "darkness" in the negative prompt while requesting a "moody night scene" in the positive prompt produces unpredictable color grading and lighting conflicts.

Balancing Prompt Weights and Guidance Scale

Fine-tuning prompt influence involves two complementary layers: the global Guidance Scale slider and individual in-prompt keyword weighting.

The Guidance Scale (CFG) Slider

The Guidance Scale slider controls the overall strength with which the Aperture model clings to the user's text instructions:

• Low Guidance (Values 3–6): Grants the model high creative latitude. Images tend to exhibit natural transitions, soft lighting, and subtle artistic textures, but the model may ignore secondary prompt descriptors.
• Balanced Guidance (Values 7–9): The optimal baseline for most commercial and concept art workflows. The model adheres reliably to complex descriptive prompts while maintaining coherent lighting, natural skin tones, and sharp edges.
• High Guidance (Values 10–14): Enforces rigid compliance with every keyword. Useful when specific colors, objects, or arrangements must appear, but runs the risk of oversaturating colors and sharpening edges excessively.
• Extreme Guidance (Values 15+): Often results in "burnt" or "deep-fried" artifacts, harsh banding, high-contrast halos around edges, and degraded textural details.

In-Prompt Keyword Weighting Syntax

When a prompt contains multiple competing elements, standard Stable Diffusion syntax allows users to adjust the emphasis of individual phrases:

• Explicit Numerical Multipliers: Encapsulate the keyword or phrase in parentheses followed by a colon and the desired numerical weight, such as (cinematic lighting:1.2) or (vibrant neon colors:1.3). A value of 1.0 represents baseline emphasis, values above 1.0 amplify the concept, and values below 1.0 (e.g., (vintage grain:0.8)) attenuate its visual prominence.
• Bracket Weighting: Wrapping terms in nested parentheses ((keyword)) provides stepped emphasis increases (equivalent to multiplying by approximately 1.1 per set of parentheses). Conversely, square brackets [keyword] reduce emphasis (equivalent to multiplying by approximately 0.9).
• Managing Concept Competition: When blending two distinct styles—such as combining architectural realism with futuristic cyberpunk aesthetics—setting (brutalist concrete architecture:1.2) alongside (subtle cyberpunk accents:0.7) prevents the neon lighting from overwhelming the geometric masonry of the building.

Step-by-Step Workflow: From Search Inspiration to Custom Generation

Lexica's unique advantage lies in the integration between its search database and generation workspace. Creators can reverse-engineer successful community prompts to calibrate their own settings:

1. Analyze Proven Prompts: Search Lexica for the intended subject or visual style. Click on an existing image to inspect its metadata panel, which displays the exact prompt, negative prompt, guidance scale, dimensions, and seed utilized.
2. Transfer to the Generate Editor: Click "Open in editor" or copy the prompt string into the Generate tab.
3. Set the Aspect Ratio First: Choose dimensions corresponding to the required output channel (e.g., 640x1024 for mobile display or 1024x640 for widescreen banners). Setting dimensions early ensures the composition develops appropriately from the initial generation.
4. Calibrate Guidance Scale: Set the Guidance Scale to 7 or 8 as a baseline.
5. Refine with Targeted Negative Prompts: Add standard hygiene terms (blurry, deformed hands, watermark) and observe the generated previews.
6. Iterate with Prompt Weights: If key elements fail to render prominently or secondary styles dominate the composition, apply numerical weights (term:1.2) to guide the model's focus without raising the overall Guidance Scale.

Generation Parameter Reference Guide

The following table summarizes the recommended configurations and typical operational boundaries for key generation parameters on Lexica:

By approaching Lexica's parameters as a coordinated system rather than isolated sliders, creators can consistently produce controlled, high-fidelity images tailored to specific editorial, commercial, and design requirements.