I Tested 10 Image Generation Models - For my SaaS Product

Image Generation Model Comparison for Thumbnail Creation

Key Concepts:

• Image-to-Image Generation: Creating new images based on an input image and a text prompt.
• Prompt Engineering: Crafting effective text prompts to guide image generation models.
• SaaS (Software as a Service): A software distribution model where applications are hosted by a provider and made available to customers over the internet.
• API (Application Programming Interface): A set of rules and specifications that allow different software applications to communicate with each other.
• Aspect Ratio: The proportional relationship between an image's width and height.
• Wave.ai: A platform offering access to numerous image generation models.
• Vibe Coding: A method of software development utilizing AI assistance for code generation.

---

I. Project Overview: Automated Thumbnail Generation SaaS

The creator is developing a SaaS product – an automated thumbnail generator – designed to simplify the thumbnail creation process for YouTubers. The project aims to replicate the creator’s own successful thumbnail creation workflow and automate it through AI. The product will be offered via a subscription model (estimated $5-$10). The core of the project relies on identifying the optimal image generation model for consistently producing high-quality, engaging thumbnails. The entire development and deployment process will be documented in a free YouTube series.

II. Methodology: Testing 10 Image Generation Models

The creator tested ten different image generation models using a consistent input image and prompt. The goal was to evaluate each model’s ability to preserve facial features, blend images effectively, generate realistic text, and produce visually appealing thumbnails. The testing was conducted through the Waves.ai platform, which provides access to a wide range of models. Approximately 20-30 images were generated per model, consuming $5 worth of credits on the platform.

III. Model-by-Model Performance Analysis

The following details the performance of each model, including cost per image, key observations, and overall suitability for thumbnail generation:

• Flux 2 Max Edit: ($0.07/image - 14 images/$1) Good face preservation but poor image blending and text quality. Potentially improvable with better prompting.
• Flux 2 Pro: ($0.06/image - 16 images/$1) Significantly worse than Flux 2 Max Edit, with distorted faces, unnatural car rendering, and blurry text. Strongly discouraged.
• Grock: ($0.022/image - 45 images/$1) Very cheap, good text generation, but poor face preservation and unexpected aspect ratio cropping.
• Alibaba WAN 2.6: ($0.035/image - 28 images/$1) Poor preservation, unnatural car rendering, but generally natural-looking image generation. Anomalous results were observed with one prompt.
• Cling Image 01: ($0.028/image - 35 images/$1) Poor image quality, barely visible car, and slow generation time (50 seconds).
• OpenAI GPT Image 1.5: ($0.053/image - 19 images/$1, $0.032/image - 29 images/$1 with aspect ratio change) Good blending and image quality, but struggles with maintaining the correct aspect ratio for YouTube thumbnails, requiring additional cropping or resizing. Cost-effective.
• Seedream 4 (Bite Studio/TikTok): ($0.027/image - 37 images/$1) Poor face preservation, but decent background and car rendering. Fast generation time (26 seconds).
• Seedream 4.5 (Bite Studio/TikTok): ($0.04/image - 25 images/$1) Improved quality over Seedream 4, but lacks creativity for YouTube thumbnails. Potential for use in editing specific image elements. Fast generation time (29 seconds).
• Google Nano Banana Pro Edit: ($0.07/image - 7 images/$1) Excellent face preservation, good car rendering, and YouTube-appropriate aesthetic. Expensive and slow (62 seconds).
• Google Nano Banana Pro Edit Ultra: ($0.06/image - 6 images/$1) The best overall performer. Consistently high-quality results, realistic images, and the ability to specify YouTube resolution. Expensive but potentially worthwhile for consistent thumbnail quality.

IV. API Key Discrepancy & OpenAI Results

A notable issue was observed with the OpenAI API key. Using the API key directly resulted in significantly lower-quality face generation compared to using the same model through Waves.ai or ChatGPT. This highlights potential inconsistencies and challenges when working directly with APIs. The creator noted that the face generation quality within ChatGPT itself was also very good.

V. Development Process & Technology Stack

The creator is documenting the entire development process using "Vibe Coding," a method leveraging AI assistance for code generation. The project will involve:

• Folder structure design
• Technology selection
• GitHub integration
• Hosting platform selection
• Authentication and payment method implementation

VI. Future Considerations & Synthesis

The creator is currently evaluating whether to rely solely on Google Nano Banana Pro Edit Ultra or combine Seedream 4.5 with a separate face-swapping model to reduce costs. The decision will depend on the desired level of consistency and the acceptable trade-off between cost and quality. The final thumbnail for this video was generated using one of these models. The creator encourages viewers to engage with the upcoming YouTube series to learn more about the project’s development and implementation.

Notable Quote:

“If you can only generate six thumbnails in $1, I mean that's good enough as well, right?” – Creator, reflecting on the cost-benefit analysis of using Google Nano Banana Pro Edit Ultra.