


0
4
This is TLB - Blind Box Style
Designed by @femboyshouta , alternate costumes for characters involved in When Reality Breaks -- imagined as K-Pop Boyband, they named it "최후소년단 — CHOEHU SONYEONDAN" or "The Last Boys", that's where TLB comes from.
They look really cool, so I decided to do this while also trying to benchmark GPT Image 2.5!
Okay, actually I performed this series of gen twice. But my first gen has problems with the prompt in terms of positioning. When I re-generated it with the updated prompt, the results where... of course, different. I keep my position on zero-shot prompt is the best way to gen with GPT Image 2 (and 2.5, apparently).
Alright, let's get to it.
The main focus of this series of generation is to see if it can handle a realistic, stylized image of an action figure well. The parameters to look at are
- Prompt-adherence of subjects and location (spatial awareness);
- Prompt-adherence of details;
- Prompt-adherence of lighting;
- Final generation quality.
For the first point, each model is doing well, no problems about it.
Second point, in terms of details, you'll see that Flare is definitely the simplest out of three, followed by GPT Image 2, and finally Sunburst. The problem with details is whether it is accurate with your prompt. Sunburst absolutely fails in that regard that it adds more context than prompted. GPT Image 2 suffers the same, with adding names on the side of the box. Accessories accuracy is a bit inaccurate with GPT Image 2 and Flare.
Third point, Flare and Sunburst are overly vibrant. GPT Image 2 is still better to get a more "natural-looking" lighting.
Fourth point, I can't emphasize this enough -- both Flare and Sunburst have that "noisy" texture of which Flare suffers the most. GPT Image 2 remains the leader in terms of noise-less image.
Final rank would be...
GPT Image 2 -> Sunburst -> Flare.
That's it for this one.
Thanks again, @femboyshouta for your designs! Absolutely fantastic!
Sampling Method
Sampling Steps
26
CFG Scale
7
Seed
600714710
Canvas Size
2560x1440

