12 image models, six influencer prompts, one surprising winner.
GPT Image 2.5 has been the default for closed-source images. So twelve frontier models got the same six AI influencer prompts, two shots each, and the one that best nailed the candid look wasn't GPT.
Image models feel like they've slowed down lately, and it's no longer obvious which one is best. The lineup:
- OpenAI / Google: GPT Image 2.5, Nano Banana 2.1, Nano Banana Pro
- Others: FLUX 3, Seedream 5.0 Pro, Grok Imagine 2.0, Ideogram 4.5, Recraft V4.1, Kling O1 Image
- Higgsfield in-house: Soul 2.0, Soul Cinema 1.5, Cinematic Studio 2.5
01The setup
Higgsfield's Supercomputer (an agent running Opus 5.5) sent each prompt to all twelve models, two images apiece, then built an HTML comparison page to browse the results without switching tools.
The six prompts are the kind of content AI influencer accounts actually post: an accidental mirror selfie, a Tokyo underground car meet, girls' night out in a restaurant booth (several people and hands), a London rainstorm (umbrella, movement), a 2006 digital-camera house party, and a UGC serum unboxing with a logo.
02Mirror selfie
Most models handled it. GPT Image 2.5 looked good at 4K, Nano Banana Pro looked very natural, FLUX 3 was realistic, and Cinematic Studio 2.5 was a standout. Nano Banana 2.1 blurred the skin, Seedream 5.0 smoothed it, and Soul Cinema mangled a hand. Grok Imagine 2.0 gave the classic AI face, plus feet growing out of the counter.
03Tokyo car meet
Nearly everyone got the cars right. GPT Image 2.5 was the best by a distance, down to a convincing hoodie in the crowd. Nano Banana Pro was close but with less skin detail; FLUX 3 nailed the reflections; Seedream had great low-fi charm but warped background faces. Recraft has a cool, artistic style but it's inconsistent.
04Restaurant booth
Multiple people is where models slip. GPT Image 2.5's food looked real, but a knife was off. Nano Banana Pro came out blown-out and too smooth, FLUX 3 too smooth. Higgsfield Soul 2.0 won with a filter that looks like what's actually trending, plus a little blur.
05London rainstorm
GPT Image 2.5 showed a noise problem here, a grain pattern most visible in the rain, and it has been getting worse. FLUX 3 made a great water splash, but the image wasn't fooling anyone. Ideogram gave her a comically tiny umbrella. Soul 2.0 again looked the most like a real candid photo, without being prompted for that style.
062006 house party
GPT Image 2.5 did well on details like the photos on the fridge and product text. Nano Banana Pro reads as AI at a glance, maybe because the internet is now full of it. FLUX 3 produced a Cheeto-like tongue and a haunted hand coming through a door. Grok's was decent but over-blurred the background, a classic giveaway. Soul 2.0 nailed it.
07UGC unboxing
GPT Image 2.5 rendered the label text cleanly. Nano Banana Pro is good at UGC but still has that uncanny, plasticky hair. FLUX 3 added a mystery hand. Soul 2.0 botched the text but looked like a frame grabbed from a real video. Soul Cinema 1.5 had the best skin texture of the whole test.
The verdict
| MODEL | TAKEAWAY |
|---|---|
| Higgsfield Soul 2.0 | Best candid influencer look. Won three of six. Not the sharpest, and weak on text. |
| GPT Image 2.5 | Most detailed, best for cars and text, but a growing noise problem. |
| Soul Cinema 1.5 | Best skin texture, inconsistent hands. |
| Nano Banana Pro | Capable, but increasingly recognizable as AI. |
| FLUX 3, Seedream 5.0, Recraft | Interesting moments, unreliable anatomy and backgrounds. |
| Nano Banana 2.1, Grok, Kling O1 | Skip. Better frontier models are cheap enough now. |
The bigger takeaway: frontier image models have slowed down, and open models are catching up. For this look, a local model like Qwen Image or the new MOSS Image is now a real alternative. Members compare model picks and prompts at smallzero.io.
$ ls ../related
$want the datasets, workflows and LoRAs behind posts like this?
they're shared inside the skool community. come join us.