Image Battle #1: Antique Brass Pocket Watch
This is our first image-model battle — a separate track from the video-model library. Same still prompt, five text-to-image models, one generation each. Every output below was produced through the fal.ai API; nothing was hand-picked from a playground gallery.
The brief is deliberately picky: macro brass, warm lamp, dust in the beam, shallow depth of field. It is the reference still we plan to feed into Veo 3.1 image-to-video next — so composition, watch type, and where the eye lands all matter.
We ran exactly one attempt per model. On first sight we did not spot hard errors or obvious glitches in any frame — the differences are about taste, attention, and value at scale. Scroll to the end of this page for all five results; the rankings and model notes come first so you know what to look for.
The prompt
Extreme close-up photograph of a hand holding an ornate antique brass pocket watch under a single warm desk lamp in a dim study at night. Intricate hand-engraved floral patterns on the watch case catch and scatter warm lamplight, fine dust motes visible in the light beam, shallow depth of field with softly blurred bookshelves in the background. Cinematic macro photography, rich shadow detail, warm color grade, photorealistic, 16:9 composition, no text, no watermarks.
All 5 images generated via fal.ai text-to-image API — one attempt per model, no rerolls.
What we did
- ▸Pipeline: text-to-image only. Platform: fal.ai for all five models.
- ▸Models tested: Nano Banana 2, Nano Banana Pro, FLUX.2 Pro, Seedream 5.0 Lite, Seedream 5.0 Pro.
- ▸Settings: 16:9, seed 42 where the endpoint supports it, one image per model.
- ▸Attempts: 1 per model — no cherry-picking, no rerolls.
- ▸First-pass check: no broken hands, no melted metal, no obvious API failures on any output.
Final ranking
- 1
Seedream 5.0 Pro
Some lighting issues if you look closely, but most viewers will not notice. Looks natural overall — best balance of quality and price at volume ($0.0675/run).
€0.06·$0.0675 USD·1 attempt - 2
Nano Banana Pro
Best single image to my eye — great attention to hand and watch — but at ~$0.15/run it is 3–4× pricier than the value tier when you scale.
€0.14·$0.15 USD·1 attempt - 3
Nano Banana 2
Strong hand detail (including realistic skin imperfections) but a different open-dial watch type than the others; the hand pulled attention away from the case.
€0.07·$0.08 USD·1 attempt - 4
FLUX.2 Pro
Beautiful watch styling, but the grip feels too straight and stiff — not how someone naturally holds a pocket watch.
€0.05·$0.05 USD·1 attempt - 5
Seedream 5.0 Lite
Watch too small in frame; dust feels localized around the watch instead of spread through the room.
€0.03·$0.035 USD·1 attempt
Model-by-model notes
Detailed impressions below — all five outputs are at the end of this page.
Seedream 5.0 Pro
ByteDance · bytedance/seedream/v5/pro/text-to-image
Dust looks natural to me — spread through the light beam the way you would expect in a dim study.
This is the model that took the risk of showing the watch face. AI image models famously struggle with time that is not ~10:10; this frame still invites you to look inside, which is a pleasant surprise.
Previous models in this test mostly kept the dial covered — possibly a safety choice baked in by developers. Here you actually want to see what is inside.
Caveat: I do not think the shadowing is fully correct. The watch sits far from the lamp but reads a bit too bright and visible for that distance. A trained eye may notice; many viewers will not.
Winner on value at scale: not the absolute best frame in isolation, but the smartest pick when you multiply cost across hundreds of stills.
Nano Banana Pro
Google · fal-ai/nano-banana-pro
My favourite image on pure quality. The hand still draws some attention — more than ideal for a product-style macro — but the way the watch is held feels natural and believable.
Dust in the beam looks normal and realistic, same as the better tiers in this test.
If budget is unlimited and you need one hero still, this is the frame I would shortlist first.
Pricing is the knock: about $0.15 per run on fal.ai — roughly 3–4× the Lite / FLUX / Seedream Pro band. Hard to justify at production volume.
Nano Banana 2
Google · fal-ai/nano-banana-2
Excellent detail on the hand — including small defects and texture that real skin has. That realism is genuinely impressive.
The trade-off: my eye went to the hand before the watch. For a watch-led brief, that is the wrong hierarchy.
Watch type differs from the other models: an open dial / hunter-style antique without the outer cover. That is a valid antique form, but my preference is not to show the dial (ciferník) directly unless the brief asks for it — possibly my own prompt discipline issue rather than a model fault.
Dust looks normal and realistic. Like Seedream Pro, it showed a readable time — models often default to 10:10 when they expose a face.
Strong mid-tier output at $0.08; just not the composition I would feed to video.
FLUX.2 Pro
Black Forest Labs · fal-ai/flux-2-pro
The watch itself looks great — case engraving and brass read well.
Holding pose feels wrong: the hand presents the watch too straight-on, almost flat to camera, without the slight angle you expect when someone cradles a pocket watch.
Dust looks normal. At $0.05 this is still a credible budget option if you only need case beauty and will crop the hand.
Seedream 5.0 Lite
ByteDance · fal-ai/bytedance/seedream/v5/lite/text-to-image
The watch reads too small — it swallowed the composition and pulled attention for the wrong reason.
Dust does not feel natural: it looks like particles emit from the watch instead of floating through the room in the lamp beam.
Cheapest run in the test at $0.035, but I would reroll or step up to Pro for anything client-facing.
Next step in the pipeline: take the ranked still (Seedream 5.0 Pro for value, Nano Banana Pro for hero quality) into Veo 3.1 image-to-video — Lite, Fast, and Standard tiers — with a separate video prompt.
Raw files, per-model cost.txt records, and the exact prompt live under public/battles/image-battles/antique-brass-pocket-watch-t2i/.
Results
All five outputs from fal.ai — same prompt, one attempt each. Click any frame for full resolution.
Raw battle archive
Side-by-side outputs, cost.txt per folder, prompt file on disk.