600+ AI Image Tests: OpenAI = Creative, Gemini = Realistic, Seedream = Fast

LateNiteSoft compared OpenAI, Gemini, and Seedream across 600+ real-world image edits, measuring quality, behavior, and speed. OpenAI is strongest for creative, transformative tasks; Gemini is best for faithful, realistic edits; Seedream is fast and often solid, with standout styles like low-poly and ukiyo‑e. They built a credit-based billing system and are prototyping a prompt classifier to auto-select models by edit type.

Key Points

• They tested 600+ generations across OpenAI gpt-image-1, Gemini 2.5 Flash Image, and Seedream 4-0-250828 using simple prompts on common photo subjects.
• Generation speed: OpenAI High ~80s (Medium ~36s), Gemini ~11s, Seedream ~9s; speeds were consistent across prompts and images.
• Gemini best preserves realism and reduces hallucinations but can be overly conservative or refuse edits, especially with human subjects.
• OpenAI excels at creative, fully transformative edits and style transfer but often re-synthesizes details, making it weak for precise tasks like background removal.
• Seedream is a fast, cost-effective middle ground with standout results in certain styles (e.g., ukiyo‑e, low poly, bokeh) but variable consistency.

Sentiment

The overall sentiment of the Hacker News discussion is largely corroborative of the article's technical findings regarding the distinct quirks and strengths of each model. Commenters generally agreed with the characterizations of OpenAI's tendency to re-synthesize images and Gemini's conservatism. However, there was a mixed and often critical view on the broader implications of AI for artists, with some expressing strong skepticism about AI art's current utility and concerns about job displacement, while others highlighted AI's potential as a productivity tool for skilled humans. The discussion, therefore, reflects a cautious but engaged stance, with specific points of agreement on model behavior and more varied opinions on the larger societal and artistic impact.

In Agreement

• OpenAI models (DALL-E 3 / gpt-image-1) tend to significantly alter faces, smooth details, re-synthesize images, and introduce a characteristic yellow/orange tint, consistent with the article's 'AI slop' observation and difficulty with preserving original details.
• Gemini is often overly conservative, prone to refusing edits (especially on people), and can return images largely unchanged, sometimes despite its interface claiming success. This aligns with the article's finding of Gemini's conservatism and literal interpretation.
• OpenAI's image generation process seems to involve deriving semantic details and then regenerating an entirely new image based on that 'description', rather than directly modifying the input, which explains its tendency to alter features and introduce 'AI slop'.
• No single AI model currently dominates all use cases; each has distinct strengths and weaknesses, making selection dependent on the specific creative goal (e.g., transformative vs. faithful edits).
• AI models are seen as a new generation of tools for accelerating productivity, but current implementations often lack sophisticated integration into existing artistic workflows.

Opposed

• One commenter strongly disagreed with the article's overall assessment, finding OpenAI results 'head-and-shoulders above the others' for artistic quality, despite acknowledging some specific failures, directly challenging the article's nuanced model comparison.
• Specific examples were provided where OpenAI failed to perform basic prompt instructions (e.g., remove background, apply fish-eye, strong bokeh blur), contradicting the notion that it is consistently 'best for transformative/creative work'.
• While the article notes Gemini's conservatism, some users found that more specific and detailed prompting could improve its cooperativeness, suggesting the model's limitations might be mitigated by user skill.
• The utility of testing model capabilities with descriptive words was questioned, suggesting that the comparison might not provide truly repeatable or informative insights into the underlying model performance.
• Regarding the broader impact on artists, several commenters strongly argued against the prediction that AI would make artists obsolete, emphasizing human creativity and the generation of new styles, contrasting with predictions of widespread job redundancy for illustrators/designers.