AI Image Generation: 5 Models, 26 Real Results
We tested five AI image models—Nano Banana 2 (Gemini 3.1 Flash), MiniMax image-01, Grok Aurora, CogView-4, and GPT (DALL-E 3)—across 26 images in two real production workflows for tabiji: 1970s Kodachrome film slides of Kyoto (18 images) and candid iPhone tourist shots of Verona for Instagram Reels (8 images).
Across 26 production images and 5 models, prompt adherence—not raw visual polish—separated usable tools from toys: Nano Banana 2 won overall at ~$0.02/image.
- Vintage film & crowd scenes (Nano Banana 2) — Scored 8.8/10 on 1970s Kodachrome Kyoto and 9/10 on Verona crowd shots, rendering legible Japanese kanji and real Kodachrome rebate markings at ~$0.02/image.
- Close-up portraits (MiniMax image-01) — Scored 8/10 on candid café portraits with the most convincing human faces (~$0.04/image), but ignored monochrome constraints (2/10 on B&W).
- All-rounder vs. laggards — Grok Aurora (~$0.07/image) finished a close second across Verona with 9/10 emotional storytelling; CogView-4 (2/10 prompt adherence) and GPT/DALL-E 3 (4–6/10, garbled text and plastic skin) finished last.
Why This Test Matters
Most AI image benchmarks test forgiving prompts like "a cat wearing a hat" or "futuristic cityscape." That tells you which model makes pretty default wallpapers—not which model follows a production brief.
Our Instagram Reels require two unforgiving aesthetics: 1970s Kodachrome slides with period-accurate grain and clipped dynamic range, and handheld iPhone tourist shots with legible street signage and imperfect framing. If a model ignores stylistic constraints on vintage film or monochrome silver gelatin prints, it will ignore your constraints on any production workflow.
The 5 Models We Tested
| Model | Provider | Cost/Image | Max Resolution | Tests |
|---|---|---|---|---|
| Nano Banana 2 | Google (Gemini 3.1 Flash) | ~$0.02 | Up to 4K | Kyoto + Verona |
| MiniMax image-01 | MiniMax | ~$0.04 | Up to 2K | Kyoto + Verona |
| CogView-4 | Z.AI (Zhipu AI) | ~$0.02 | 720×1440 | Kyoto only |
| Grok Aurora | xAI | ~$0.07 | Up to 2K | Verona only |
| GPT (DALL-E 3) | OpenAI | ~$0.04–$0.08 | 1024×1792 | Verona only |
Test 1 (Kyoto) ran Nano Banana 2, MiniMax, and CogView-4 across four landmarks (Fushimi Inari, Kinkaku-ji, Arashiyama, and Gion). Test 2 (Verona) dropped CogView-4 and added GPT and Grok Aurora across two candid iPhone prompts.
Test 1: Vintage 1970s Kyoto — Can AI Fake Film Photography?
We prompted all three models for 1970s Kodachrome film: warm saturated midtones, vintage grain, off-center tourist framing, and period-accurate details. Here is Fushimi Inari:
▶ Animated — Vintage POV Reel Clips
These clips show the same images after Remotion rendering — with film grain, vignette, and slow drift animation applied.
Across all four landmarks, Nano Banana 2 nailed the brief: warm Kodachrome midtones, legible kanji ("奉" and "納" carved on the torii columns), and casual snapshot framing. MiniMax had pleasing color science but ignored POV instructions and composed like a modern travel magazine. CogView-4 ignored the vintage prompt entirely—rendering HDR orange-teal grading and garbled kanji.
Kinkaku-ji and Arashiyama showed the exact same split:
▶ Animated — Kinkaku-ji Reel Clips
▶ Animated — Arashiyama Reel Clips
The Black & White Test That Broke Two Models
For Gion at dusk, we specified "absolutely no color whatsoever, silver gelatin print"—a geisha walking down a lantern-lit alley on black-and-white film.
▶ Animated — Gion Evening Reel Clips
Nano Banana 2 delivered pure monochrome with zero color bleed—inky machiya shadows, glowing lanterns, and Tri-X-style silver grain reminiscent of Daidō Moriyama. MiniMax (2/10) rendered warm amber lanterns and teal shadows. CogView-4 (0/10) output neon orange lanterns and bright red obi accents.
When we asked for "absolutely no color whatsoever," Nano Banana 2 gave us black and white. MiniMax and CogView-4 gave us whatever their priors wanted.
Test 2: Phone Camera Realism — Verona
The other half of our pipeline needs images that pass as unfiltered iPhone photos—handheld framing, natural depth of field, crowd density, and readable signs. We tested GPT (DALL-E 3), Grok (Aurora), Nano Banana 2, and MiniMax across two Verona prompts.
Which AI model handles crowd scenes best?
Prompt: An exhausted female tourist photographed from behind in a dense crowd outside Casa di Giulietta in Verona. iPhone photo quality — handheld, natural grain, depth of field, motion blur. Juliet's bronze statue visible ahead. Tourist overwhelm is palpable.
Nano Banana 2 won (9/10): legible "Juliet" souvenir text, an accurate Fjällräven Kånken logo on the backpack, and a raised phone screen in the foreground that sells the tourist crush. Grok Aurora (8.5/10) delivered the most expressive facial fatigue on the main subject, with slight face melting deeper in the crowd. MiniMax (7/10) shot it like a moody Sony A7III thriller in a dark alley, and GPT (6/10) produced porcelain skin and gibberish signage ("SLOLEROCVEE").
Is GPT good for portrait-style travel photos?
Prompt: Two Italian men in their 50s laughing and drinking wine at an outdoor café in Piazza delle Erbe, Verona. Candid iPhone shot — natural light, movement blur, authentic body language. Amarone bottle visible. Medieval tower in background.
MiniMax won portraits (8/10): the two men look like specific, lived-in human beings with documentary lighting. Grok (7/10) nailed the candid laughter (despite matching sweaters and missing wine glasses). Nano Banana 2 (6.5/10) rendered a legible "Amarone" bottle label and medieval tower, but the pose felt staged. GPT (4/10) looked like an oversaturated stock illustration.
📱 Key Finding: Different Models Win Different Scene Types
Nano Banana 2 dominates crowd and location scenes — text rendering, brand accuracy, and iPhone aesthetic make it the clear winner for anything with signage, landmarks, or branded items.
MiniMax dominates intimate people/portrait scenes — when the faces are the subject, MiniMax's human rendering is unmatched.
Grok is the strongest all-rounder — close second in both categories, best emotional storytelling overall. If you're picking one model for everything, pick Grok.
GPT/DALL-E 3 is dead last for photorealism. Garbled text, over-rendered aesthetics, and uncanny faces make it unsuitable for content that needs to pass as real.
Prompt Engineering Makes or Breaks It
After the first Kyoto pass, we rewrote our prompts with explicit optical and film-stock physics:
- V1 (vague): "Shot on Kodachrome film, vintage feel"
- V2 (specific): "Warm saturated midtones, slightly cool shadows, limited dynamic range with clipped highlights and blocked shadows. Include scan artifacts: dust specks, hair, scratches. Amateur composition, slightly off-center. Chromatic aberration, lens softness at edges. No borders, no frame edges."
With V2, Nano Banana 2 generated authentic Kodachrome film rebate markings ("12 KODACHROME" in orange-on-black along the border with frame arrows) and exposure-dependent grain that clumps in the shadows like real silver halide crystals.
MiniMax shifted warmer and added a subtle light leak, but ignored the physical scan artifacts and film stock details.
Detailed prompt engineering only pays off on models that actually listen. Nano Banana 2's prompt ceiling is 10/10; MiniMax improves marginally; CogView-4 ignores technical specs.
Combined Scorecard: All 5 Models
| Category | Nano Banana 2 | MiniMax | CogView-4 | Grok Aurora | GPT (DALL-E 3) |
|---|---|---|---|---|---|
| Prompt Adherence | 9.5/10 | 5/10 | 2/10 | 7/10 | 5/10 |
| Vintage Film | 8.8/10 | 5.9/10 | 3.6/10 | — | — |
| Crowd Scenes | 9/10 | 7/10 | — | 8.5/10 | 6/10 |
| People/Portraits | 6.5/10 | 8/10 | — | 7/10 | 4/10 |
| Text / Kanji Accuracy | 8/10 | 4/10 | 1.5/10 | 6/10 | 3/10 |
| Prompt Eng. Ceiling | 10/10 | 5/10 | 3/10 | 7/10 | 5/10 |
| Emotional Storytelling | 7/10 | 7/10 | 6/10 | 9/10 | 5/10 |
| Cost per Image | ~$0.02 | ~$0.04 | ~$0.02 | ~$0.07 | ~$0.04–$0.08 |
What to Use for What
- Crowd scenes, landmarks, signage, or CJK text → Nano Banana 2 (~$0.02/image). Unbeatable prompt adherence and legible text.
- Vintage film / strict stylistic constraints → Nano Banana 2. The only model that respects monochrome and film-stock physics.
- Close-up portraits & human faces → MiniMax image-01 (~$0.04/image). Best skin texture and natural lighting.
- Single all-rounder model → Grok Aurora (~$0.07/image). Strong second place everywhere and 9/10 emotional storytelling.
- Skip for photorealism → GPT/DALL-E 3 (plastic skin, garbled text) and CogView-4 (ignores style instructions).
Pricing Comparison
| Factor | Nano Banana 2 | MiniMax | CogView-4 | Grok Aurora | GPT (DALL-E 3) |
|---|---|---|---|---|---|
| Cost per image | ~$0.02 | ~$0.04 | ~$0.02 | ~$0.07 | ~$0.04–$0.08 |
| 6-image Reel cost | ~$0.12 | ~$0.24 | ~$0.12 | ~$0.42 | ~$0.24–$0.48 |
| Max resolution | Up to 4K | Up to 2K | 720×1440 | Up to 2K | 1024×1792 |
| Free tier | Yes (Gemini API) | Limited | Limited | Limited | No |
| API complexity | Moderate (SDK) | Simple REST | Simple REST | Simple REST | Simple REST |
| Latency | ~8–15s | ~10–20s | ~15–25s | ~10–15s | ~10–20s |
At ~$0.12 per 6-image Reel (plus a free tier on Google AI Studio), Nano Banana 2 is both the cheapest and highest-scoring model for location and vintage workflows.
How to Use These Models (Code Examples)
Here are the minimal API calls for all five models:
Nano Banana 2 (Google Gemini 3.1 Flash Image)
import google.generativeai as genai
genai.configure(api_key="YOUR_GEMINI_API_KEY")
model = genai.GenerativeModel("gemini-3.1-flash-image-preview")
response = model.generate_content(
"A vintage 1970s Kodachrome photograph of Fushimi Inari torii gates. "
"Warm saturated midtones, slightly cool shadows, limited dynamic range. "
"Include scan artifacts: dust specks, scratches. Amateur composition.",
generation_config=genai.GenerationConfig(
response_modalities=["IMAGE", "TEXT"]
)
)
# Save the image
for part in response.candidates[0].content.parts:
if part.inline_data:
with open("output.png", "wb") as f:
f.write(part.inline_data.data)
MiniMax image-01
curl -X POST "https://api.minimax.chat/v1/image_generation" \
-H "Authorization: Bearer YOUR_MINIMAX_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "image-01",
"prompt": "A vintage 1970s photograph of Fushimi Inari torii gates...",
"aspect_ratio": "9:16",
"response_format": "url"
}'
CogView-4 (Z.AI)
curl -X POST "https://open.bigmodel.cn/api/paas/v4/images/generations" \
-H "Authorization: Bearer YOUR_ZAI_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "cogview-4-250304",
"prompt": "A vintage 1970s photograph of Fushimi Inari torii gates...",
"size": "720x1440"
}'
Grok (xAI Aurora)
curl -X POST "https://api.x.ai/v1/images/generations" \
-H "Authorization: Bearer YOUR_XAI_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "grok-2-image-1212",
"prompt": "Phone camera photo of exhausted tourist in crowd...",
"n": 1,
"response_format": "url"
}'
GPT / DALL-E 3 (OpenAI)
curl -X POST "https://api.openai.com/v1/images/generations" \
-H "Authorization: Bearer YOUR_OPENAI_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "dall-e-3",
"prompt": "Phone camera photo of exhausted tourist in crowd...",
"n": 1,
"size": "1024x1792",
"quality": "standard"
}'
Full Reels: Side by Side
Here are the complete 21-second Vintage POV Reels assembled from the Kyoto outputs—identical Remotion timing, vignette, slow drift, and text overlays, differing only in the source images:
Both Reels use identical Remotion compositions and timing. Nano Banana 2's Kodachrome color science locks right into the animated film-grain layer, whereas MiniMax's modern rendering clashes with the vintage overlay.
All images in this comparison were generated from identical prompts on the same day (March 10, 2026). No post-processing was applied. The images shown are direct outputs from each model's API.
Newsletter
Get the next post by email.
One email when I publish something new. No spam, no fixed schedule, unsubscribe anytime.