[Firefly Study Notes] Day 12 - That Person I Often See
Good morning.
It has already been a week since last Friday when I hit a dead end at my main job; time flies.
Anticipating that such a situation might happen again today, I am giving this a try starting from the morning.
Today, I intended to conduct yesterday's verification using "human images."
So, when I generated an output using the simple prompt for a woman I had prepared...
The woman with the short haircut that I often see on Adobe Stock appeared!!
✅ The level of perfection as a photograph is extremely high
❌ But, the second woman from the left... I clearly have a sense of having seen her before!
I thought "Huh?" when she appeared during the test, so this doesn't seem to be a coincidence.
So, I decided to dig deeper into this matter.
✅ Why does it become "someone I've seen before"?
🔹 Firefly is trained on Adobe Stock
Adobe Firefly is said to be trained based on "Adobe Stock" assets as copyright-clean image data.
Because of this, it seems that human figures that frequently appear in stock photos (short hair, gentle smiles, natural light-style compositions, etc.) have become templated.
🔍 What actually happened
As mentioned above, even though I was generating images with Firefly, a situation occurred where facial features close to "sample images often seen on Adobe Stock" appeared frequently.
💡 How to deal with this?
✅ Keep the prompt as is, but avoid the template face:
A portrait photo of a young Japanese woman in her twenties with short dark brown hair and natural makeup, slightly asymmetrical features, candid expression, white background, wearing a gray sleeveless dress
Add "slightly asymmetrical features" and "candid expression" to make it look less like stock.
Adjusting only the texture, such as using "photorealistic" or "soft portrait light," also seems effective.
Although it deviated from what I originally wanted to do, it was a good realization.
Next time, I would like to carry out the 'verification of composition and style using portrait images'.