Parameter Tuning Guide for Text-to-Image Models

This document will introduce how to quickly adjust and optimize the parameters and configurations of text-to-image models to achieve the best generation results.

1. AnimeSharp (Anime Super-Resolution Model)

Core Positioning

A super-resolution model specifically optimized for anime styles. It excels at preserving line sharpness and color saturation, outperforming general super-resolution models (such as Real-ESRGAN) by 40% in anime scenarios.

Parameter Settings (Must-Tune Items)

Parameter Name	Value Range	Description	Recommended Setting (Default)
Magnification	2x/4x	Controls output image size; 4x is the model-optimized multiple with best results	4x
Sharpening Strength	0.1-1.0	Adjusts sharpness of anime lines. Too high may cause noise; too low makes lines blurry	0.7 (balances sharpness and naturalness)
Denoise Threshold	0-50	Removes noise from original images (e.g., scan spots). Higher values mean stronger denoising but potential detail loss	20 (suitable for moderately noisy images)

Prompt Engineering (Input Instructions)

• Essential Elements: Clearly mark "anime style" + image subject (e.g., "2D anime girl illustration," "Japanese anime scene") to prevent the model from misclassifying it as realistic style.
• Example:
  Input image + supplementary description "anime-style character portrait with clear lines and high-contrast colors." The model will prioritize optimizing key anime details such as hair strands and clothing textures.

Applicable Scenarios

Anime illustration enlargement, old anime screenshot restoration, game concept art HD enhancement. Not recommended for realistic photos (may cause distortion of faces/scenery).

2. FLUX.1-schnell (Fast High-Quality Text-to-Image)

Core Positioning

Adopts latent adversarial diffusion distillation technology, capable of generating high-definition images in 1-4 steps. It balances speed and quality, supporting a maximum resolution of 2048×2048.

Parameter Settings (Core Optimization Items)

Parameter Name	Value Range	Description	Scenario Adaptation Strategy
Generation Steps	1-4 steps	More steps mean richer details but longer processing time: 1 step (fast but fewer details), 4 steps (best quality)	Use 1 step for quick previews, 3-4 steps for final output
Resolution	1024×1024/1536×1536/2048×2048	High resolutions require asynchronous interface (to avoid timeouts), suitable for poster/wallpaper creation	1024×1024 for social media, 2048×2048 for printing
Style Strength	0.5-2.0	Controls style transfer intensity. Higher values mean closer adherence to prompt descriptions (e.g., more prominent "cyberpunk" features)	1.2 (balances fidelity and creativity)

Prompt Engineering (Structured Template)

Subject Description (required): [Specific character/scene, e.g., "silver-haired cat-eared girl," "futuristic city night view"]  
Style Specification (required): [anime/realistic/oil painting, etc., e.g., "Makoto Shinkai-style animation scene"]  
Detail Supplement (optional): [Lighting/colors/composition, e.g., "backlight effect, water reflection, warm color tone"]  
Quality Requirements (optional): [Clarity/detail level, e.g., "8K precision, individual hair strands visible"]