Краш-тест нейронных сетей для генерации фотосессий

CRASH TEST — I took and compared neural networks for generating complex photo sessions. Three references were given: face, clothing, location. The essence: with one prompt, without any fuss, to create a decent shot. Here’s what came out.

4th place — ChatGPT. Surprisingly, it drew a completely different person, meaning it just couldn’t hold the face. And yes, I tried different prompts and approaches, but it still didn’t manage.

3rd place — Seedream 5 Lite. It started to deviate in poses: instead of one specified pose, it made several different ones. And almost everywhere the same problem — the face looks like a sticker on top, not integrated into the scene. By the way, I wrote about this problem earlier.

2nd place — Qwen Images. Overall, quite decent: it holds the face, transfers the clothing. But for some reason, it added freckles that weren’t in the reference. Plus, it couldn’t make a vertical shot — it dragged into horizontal, apparently orienting itself on the reference location.

1st place — Nano Banana, both Pro and 2. Here it’s really cool: it retains the face, transfers the clothing neatly, and the scene doesn’t fall apart. In some attempts, the Pro version even transferred not just the suit but the entire look.

Prompt:

Use Image 1 as the exact reference for the female model’s face, hair, and features. Use Image 2 as the strict reference for the winter outfit and clothing materials she is wearing. Use Image 3 for the background environment, placing the model in the snowy Alpine mountain landscape. Low-angle full-body shot, looking up slightly at the model, making her appear heroic and powerful against the mountain peaks. The model is standing in a strong, confident heroic stance, slightly adjusting the heavy collar of the winter jacket. She has an intense gaze directly into the lens, with a serious, focused expression. Her hair is slightly wind-blown. Natural overcast winter daylight, soft light reflection from the snow acting as a fill light. Shot on Medium Format Hasselblad H6D-100c, 50mm lens, f/4 for a sharp subject and softly blurred snowy background. Highly detailed skin texture, realistic fabric physics on the winter coat. Kodak Portra 400 film look, crisp cold cinematic color grading, raw photo, 8k resolution.

In short, it ended up being a bit boring: there was no intrigue, Banana wiped everyone out in one go. But maybe you have a different experience, share it, it would be interesting to compare.