DEV Community: Evan Li

Designing a quiz result that shows its limits

Evan Li — Wed, 14 Oct 2026 02:41:21 +0000

A result page can be mathematically consistent and still invite the wrong conclusion. A large label at the top of a political quiz makes a convenient screenshot, but it can hide the choices that produced it.

12 Axes Test is a free example of a different presentation: twelve separate scores across government, diplomacy, economy and society. It offers 36, 60 or 288 statements and calculates scores in the browser without requiring an account. Here are three design questions that this kind of interface raises.

What is the primary output?

A label compresses a vector into a name. That compression loses information. Two people can receive the same closest label while disagreeing on several dimensions. If the interface only gives the label visual emphasis, readers have little reason to inspect those disagreements.

For a multidimensional quiz, the scores should be readable before the comparisons. Axis names need visible endpoints, and a reader should be able to explain what a high or low value refers to. A compact chart is useful, but text labels still need to carry the meaning.

What does a comparison claim?

A nearest match is relative to the comparison set. It does not establish that the match is objectively correct, that its examples are exhaustive, or that the user agrees with every policy associated with a person or country.

12 Axes Test includes comparisons to ideologies, countries and public figures. These profiles are editorial estimates. They are discussion aids, not scientific diagnoses or voting advice. Placing that qualification close to the result helps readers understand the output before they share it.

What happens when a result is shared?

Scoring in a browser and sharing a result are different operations. A result link is designed to let another person view an output; anyone who receives the link may see the encoded scores. A share control should make that consequence understandable. Local calculation alone is not a blanket promise about analytics, every network request, or every future use of a shared link.

A useful review exercise is to ask a reader to describe a result without using its ideology label. Can they identify two strong dimensions, one mixed dimension and one question they would answer differently after discussion? If they can, the interface has given them more than a badge.

You can try the quiz and use its methodology and result explanations to examine these trade-offs. This article introduces the product and a design perspective; it does not validate the questionnaire as a scientific assessment.

Building an AI Image Generator SaaS in 2026: My Tech Stack and Lessons

Evan Li — Tue, 05 May 2026 14:18:30 +0000

I've spent the past few months working on AI image generation tooling, and a lot of friends have asked me what stack I'd use if I were starting fresh in 2026. Here's the honest answer, with the gotchas I hit along the way.

This isn't theoretical — it's based on what I've seen working in production at Nano AI (a free Nano Banana Pro workbench) and what I'd build into the next product.

The Stack

Layer	Choice	Why
Frontend	Next.js 15 (App Router) + Tailwind	Boring, fast, server components for SEO
Auth	NextAuth	OAuth (Google/GitHub) covers 95% of users
DB	PostgreSQL (Supabase or Neon)	Free tier covers MVP, scales when needed
Image Model	Nano Banana Pro / SDXL fallback	Cost per image determines unit economics
Storage	Cloudflare R2	S3-compatible, 10x cheaper for image-heavy apps
Payment	Stripe	Don't reinvent this
Hosting	Vercel (frontend) + Cloudflare Workers (image proxy)	Edge for cold starts
Analytics	PostHog (self-hosted)	Privacy-friendly, full event control

Total monthly fixed cost at MVP scale: ~$30-50. The real cost is per-image inference.

The Critical Decision: Where Do You Run the Model?

This is the make-or-break choice for an AI image SaaS. Three options:

Option 1: Hosted API (OpenAI, Stability)

✅ Zero ops
❌ Per-image cost is high — kills free tiers
❌ Vendor lock-in

Option 2: Self-host on RunPod / Modal / Replicate

✅ Cheaper per image
❌ Cold start pain (10-30s for first request after idle)
❌ More ops

Option 3: Use a small efficient model via cheap inference provider

✅ Best per-image cost (Nano Banana Pro is in the $0.001-0.003 range)
✅ Sub-second inference
❌ Quality ceiling lower than DALL-E 3 for complex prompts

For a freemium SaaS, Option 3 is the only thing that makes the unit economics work if you offer a free tier.

I wrote about the actual cost breakdown across providers — the gap between the most expensive and cheapest option is 40x.

Free Tier Strategy

This is the part most builders get wrong. Here's what works:

Don't offer "unlimited free"

You'll get bot abuse. Within a week you'll be paying for someone else's bulk generation.

Don't offer "10 free per month"

Visitors will hit the limit on their first session, never come back.

Do offer "10 free per day"

Resets daily. Keeps engagement. Shows the product enough times to convert.

Implementation: Redis or Postgres counter keyed by IP + cookie. Reset every 24h.

The Bot Abuse Problem

Within the first week of putting up an AI image generator, you'll have:

Scrapers hitting your /api/generate endpoint
Users with disposable emails creating 100 accounts
Distributed attacks from residential proxies

Defense in depth:

Cloudflare Turnstile on signup (free, lighter than reCAPTCHA)
Rate limit by IP + device fingerprint (not just IP — proxies are cheap)
Image hash deduplication — if the same prompt+seed already ran today, return the cached image
Behavioral checks — humans take >0.5s to hit "generate"; bots don't

What I'd Skip in 2026

A few things that seemed essential a year ago but really aren't for an MVP:

Custom auth flow — NextAuth handles 95% of cases, don't waste time building from scratch
Microservices — one Next.js monolith handles 10K req/min easily on Vercel
Proprietary AI model — you don't need to fine-tune your own model in 2026; the open-weight ecosystem is good enough for most use cases
Real-time collab features — adds complexity, rarely needed for image gen apps

Lessons (the hard-won ones)

Cost dictates everything. Get the per-image cost down before adding features. Otherwise you're building a money-losing machine.
Cache aggressively. A surprising number of "AI generations" are people typing similar prompts. Cache by prompt+seed.
The model's quality ceiling won't matter if your prompt UX is bad. Build a great prompt-input experience (suggestions, examples, history) and a 7/10 model beats a 9/10 model with a bad UX.
Free tier conversion is brutal in this space. Average is ~1-2% (paid). If you're below 0.5%, look at your free-to-paid friction, not your pricing.
Image hosting is its own line item. Generating a 1024x1024 PNG is one cost. Storing/serving it for 1000 visitors is another. Use R2 or B2.

Try the Stack in Action

If you want to see one concrete instance of this stack running in production, Nano AI is exactly the thing — Next.js + Nano Banana Pro + Cloudflare R2 storage.

Try the workbench, look at the load behavior in DevTools, and you'll get a sense of what "Option 3" feels like in practice.

Closing

The interesting thing about 2026 is that you can ship a real AI product as a single founder for under $50/month in fixed costs. The bottleneck isn't infrastructure or models — it's distribution and prompt UX.

If you're building anything in this space, I'm happy to chat. Drop your project in the comments.

Nano Banana Pro vs DALL-E 3 vs Midjourney: A Practical Comparison From Someone Who Actually Uses All Three

Evan Li — Tue, 05 May 2026 14:17:29 +0000

There are roughly 50 articles on the internet titled "DALL-E vs Midjourney vs X." Most of them are written by people who tried the trial of each and stitched together opinions. This one isn't — I've been using all three almost daily for a year, and I'll try to be honest about where each one wins and loses.

The newcomer in the comparison is Nano Banana Pro, accessed through Nano AI, which I've added to my regular rotation in 2026.

TL;DR Up Front

Use case	Best for	Why
Photorealism for marketing	DALL-E 3 HD	Best at hands, faces, text-in-image
Stylized art	Midjourney	Best taste in composition, lighting
Cheap, fast, "good enough"	Nano Banana Pro	1/40th the cost of DALL-E, 60% the quality
Production SaaS embed	Nano Banana Pro	Cost-per-image is what matters at scale
Specific artist style	Midjourney v6	Strongest style control

If you want only one tool, pick based on what you're actually shipping. The "best model" is the one that fits your unit economics + quality bar.

What I Actually Use Each For

DALL-E 3 → Polished blog hero images for client work where the budget allows. Also when I need text rendered in the image (it's better at this than the others).

Midjourney → Mood boards, concept exploration, anything where I want the model's "taste" to influence the result. Best when I don't know exactly what I want.

Nano Banana Pro (via Nano AI) → Fast iteration on placeholder images, free-tier features in side projects, prompt experimentation. Also: when I just need 10 variations cheaply.

Quality Comparison: 5 Actual Prompts

I ran these 5 prompts on each tool. Honest assessment below.

1. `cinematic close-up of a fox in autumn light, photorealistic`

DALL-E 3: Excellent. Fur detail is photographic, eyes are right.
Midjourney: Excellent but more stylized. Looks like a "fox photograph by Annie Leibovitz."
Nano Banana Pro: Good. Photorealistic, slightly less detail in fur. Good enough for a blog hero.

Winner: DALL-E 3, but barely.

2. `flat illustration, retro 90s computer terminal, editorial style`

DALL-E 3: Decent but tends toward photorealistic even with "flat illustration" specified.
Midjourney: Beautiful. Best aesthetic.
Nano Banana Pro: Surprisingly clean. Pulls off "flat" better than DALL-E.

Winner: Midjourney for aesthetics, Nano Banana for following the brief.

3. `studio product shot, ceramic coffee mug on marble surface`

DALL-E 3: Excellent. Production-ready.
Midjourney: Excellent but slightly over-stylized for a product shot.
Nano Banana Pro: Excellent. For this category, the cheapest tool wins — they all produce production-ready output.

Winner: Tie. Nano Banana wins on cost.

4. `cyberpunk courier, full body, neutral pose, concept art`

DALL-E 3: Good but generic.
Midjourney: Best by a mile. Composition, lighting, character design all stronger.
Nano Banana Pro: Good but less stylish than Midjourney.

Winner: Midjourney.

5. `Tokyo back alley at night, light rain, atmospheric`

DALL-E 3: Solid, photorealistic.
Midjourney: Most cinematic.
Nano Banana Pro: Excellent. Very atmospheric. The "atmospheric" anchor word works particularly well on this model.

Winner: Midjourney for vibe, Nano Banana for unexpected quality.

Score Card

Category	DALL-E 3	Midjourney	Nano Banana Pro
Photorealism	9/10	7/10	7/10
Stylized art	6/10	9/10	7/10
Speed	6/10	6/10	9/10
Cost (per image)	3/10	4/10	10/10
Prompt fidelity	8/10	7/10	7/10
Hands/text/faces	8/10	7/10	6/10
API integration	9/10	4/10 (no public API)	8/10
Best for SaaS	⚠️ Expensive	⚠️ No API	✅

Where Nano Banana Pro Surprises

Two things genuinely surprised me:

Speed. Sub-second on 512x512. This changes the UX of any tool — from "wait for the spinner" to "instant feedback."
Cost-quality ratio for placeholder/draft work. If you don't need the image to be the final asset, the cost difference (40x cheaper) makes Nano Banana Pro the obvious default.

Where it loses: complex prompts, stylized fidelity to a specific reference, and small-detail accuracy (hands, text-in-image).

What I'd Recommend

Indie SaaS founder shipping a free tier → Start with Nano Banana Pro. You can always upgrade to DALL-E for paid tiers.
Designer making mood boards / pitch decks → Midjourney. Its "taste" is unmatched.
Marketing team making client deliverables → DALL-E 3 HD. Photorealism + text-in-image makes it production-ready.
Developer just experimenting → Try Nano AI's free workbench first. No signup, instant feedback.

Closing

The "best AI image generator" question is the wrong question. The right question is: what are you optimizing for?

If it's quality at any cost — DALL-E 3 or Midjourney.
If it's quality-per-dollar at scale — Nano Banana Pro.

I use all three depending on the job. So should you.

If you want to compare for yourself, both DALL-E and Midjourney have trial credits, and Nano Banana Pro is free to test at nanoai.run with no signup.

The Real Cost of Adding AI Image Generation to Your SaaS in 2026

Evan Li — Tue, 05 May 2026 14:16:28 +0000

Last month I priced out adding AI image generation to a side project. The numbers surprised me — not because they were high, but because of how much they varied between providers. If you're building anything that needs images (content tools, e-commerce mockups, design helpers, blog illustrators), this might save you a few hours of spreadsheet work.

The Setup

Imagine a basic free tier: 10 image generations per visitor per day. With even 1,000 daily active users, you're at 10,000 images/day = 300,000/month.

Here's what that costs across providers (all numbers approximate, current as of early 2026):

Provider	Per-image cost	Monthly @ 300K
OpenAI DALL-E 3 (Standard)	$0.040	$12,000
OpenAI DALL-E 3 (HD)	$0.080	$24,000
Stability AI (SDXL via API)	$0.009	$2,700
Replicate (SDXL)	$0.0023	$690
Self-hosted SDXL on RunPod	$0.005	$1,500 (incl. cold start cost)
Nano Banana Pro	~$0.001-$0.003	$300-900

Let those numbers sink in for a second. DALL-E 3 HD is 80x more expensive than the cheapest option. For a free tier, that's not a small detail — it's the difference between a feature that makes business sense and one that bankrupts you.

What Actually Drives the Difference

Three things:

1. Model size

DALL-E 3 is rumored to be in the multi-billion-parameter range. Stable Diffusion XL is ~3.5B. Nano Banana Pro is in the "well under 1B" range, optimized via distillation.

Smaller models = less GPU time per inference = lower per-image cost.

2. Hosting margin

When you use OpenAI's API, you pay for: model inference + their infrastructure margin + their R&D recoup + their brand. That stack is roughly 5-10x the raw compute cost.

Self-hosted or smaller-provider options strip out the brand premium.

3. Quality-cost tradeoff (the honest part)

Bigger models are still better for complex prompts, hands, text-in-image, and stylistic fidelity to specific artists.

If you're building a "make me a Pixar-style movie poster" product, you probably need DALL-E 3 or Midjourney.

If you're building a "free draft image for my blog post" feature, the smaller models are 80% as good for 1/40th the price. For most freemium SaaS use cases, that's the right tradeoff.

What I'd Do (Practical Advice)

If I'm shipping a new product with image generation:

Start with the cheapest option that meets the quality bar. Cost compounds fast at scale.
Cache aggressively. A lot of "AI image generation" can actually be "search a cache of pre-generated images" if your prompt space is constrained.
Offer the expensive option as paid tier. Pay-per-use makes the unit economics work.
Measure prompt quality, not provider quality. A great prompt on a small model often beats a lazy prompt on a big model.

Try a Cheap Model Before You Commit

If you want to feel out where the cheap end of the market actually lands quality-wise, Nano AI has a free workbench powered by Nano Banana Pro. No signup required to test.

Run a few of your real prompts there. Compare to DALL-E. You'll quickly see whether the cheaper model meets your quality bar — for me, on a content tool side project, it did.

Closing Thought

The conversation around "AI is too expensive to ship" is mostly outdated by mid-2026. The cost has come down 10-100x in 18 months. The bottleneck now is knowing which model to use for which use case — and being willing to test rather than default to the most expensive.

If you've shipped image generation in a SaaS product, I'm curious what your actual unit economics look like. Drop them in the comments.

5 Prompt Recipes That Reliably Produce Great Images on Nano Banana Pro

Evan Li — Tue, 05 May 2026 14:15:27 +0000

After spending way too many hours testing prompts on Nano Banana Pro through the Nano AI workbench, I noticed a pattern: this model rewards a specific prompt structure that's different from what you'd write for DALL-E 3 or Midjourney.

This isn't a "1000 prompts you can copy" listicle. It's 5 actual recipes I keep coming back to, with the why behind each.

Why Recipes Beat Random Prompting

Most "prompt engineering" advice is just "add more adjectives." That works on bigger models like DALL-E 3 because they have huge prompt-following capacity. Smaller models like Nano Banana Pro reward terseness — long prompts confuse them.

The recipes below all share a structure:

[subject] + [lighting/time] + [medium/style anchor]

Three slots. Short. The model fills in everything else.

Recipe 1: The "Cinematic Portrait"

cinematic close-up of [subject], golden hour, photorealistic, 35mm film

Examples:

cinematic close-up of a fox, golden hour, photorealistic, 35mm film
cinematic close-up of an elderly fisherman, golden hour, photorealistic, 35mm film

Why it works: "35mm film" is a strong style anchor that the model recognizes from photography metadata in training data. "Golden hour" forces warm lighting. "Close-up" prevents the model from getting lost in background detail.

Recipe 2: The "Flat Editorial Illustration"

flat illustration, [scene description], muted palette, editorial style

Examples:

flat illustration, retro 90s computer terminal on a wooden desk, muted palette, editorial style
flat illustration, a person reading on a balcony at dawn, muted palette, editorial style

Why it works: "Editorial illustration" is a specific genre with clean lines, limited color, and intentional composition. The model latches onto it cleanly. Try it for blog hero images.

Recipe 3: The "Studio Product Shot"

studio product shot, [object] on [surface], soft shadows, [color] background

Examples:

studio product shot, ceramic coffee mug on marble surface, soft shadows, beige background
studio product shot, vintage pocket watch on velvet, soft shadows, deep blue background

Why it works: This is the most "generation-ready" recipe — Nano Banana Pro produces near-photoreal product shots with this template that you can drop straight into a landing page.

Recipe 4: The "Mood Landscape"

[place] at [time], [weather/atmosphere], wide shot, atmospheric

Examples:

Tokyo back alley at night, light rain, wide shot, atmospheric
Mongolian steppe at dusk, gathering storm, wide shot, atmospheric

Why it works: "Wide shot" and "atmospheric" together push the model toward landscape composition. Without them, it tends to default to mid-range portrait framing.

Recipe 5: The "Character Card"

[character description], full body, neutral pose, [background], concept art

Examples:

cyberpunk courier, full body, neutral pose, gray studio background, concept art
medieval librarian, full body, neutral pose, gray studio background, concept art

Why it works: "Concept art" + "neutral pose" + "studio background" gives you a character reference sheet that's useful for further editing or as a placeholder asset.

Pattern: Anchor Words

The thing all 5 recipes share is anchor words:

cinematic, photorealistic, 35mm film
flat illustration, editorial
studio product shot, soft shadows
wide shot, atmospheric
concept art, neutral pose

Anchor words are short phrases that the model treats as style/composition triggers. They do more work than 5 adjectives.

Try Them

If you want to test these recipes, Nano AI's free workbench has no signup wall — paste any of the prompts above and you'll get an output in a few seconds.

I keep meaning to write a longer post on why the model responds to anchor words rather than long descriptions. Let me know in the comments if that's interesting.

Disclaimer: I'm an active user of Nano AI but not affiliated with the company. These recipes are from my own testing.

How Nano Banana Pro Model Is Quietly Changing AI Image Generation in 2026

Evan Li — Tue, 05 May 2026 14:08:16 +0000

If you've been keeping an eye on AI image generation, you've probably noticed the field has shifted from "wow, it can draw a cat" (2022) to "wait, that's photorealistic and cost me $0.001" (2026). Behind that shift is a quieter trend: smaller, faster, cheaper models that are surprisingly competitive with the heavyweights.

One of those models is the Nano Banana Pro Model — a compact text-to-image model that's been making the rounds for its speed-to-quality ratio. I've been using it through Nano AI (a free creation workbench) and wanted to share what makes it interesting from a developer's perspective.

What is Nano Banana Pro?

Nano Banana Pro is a text-to-image model optimized for fast inference at low compute cost. The "nano" naming hints at the design philosophy:

Fewer parameters than DALL-E 3 / SDXL — but optimized via distillation
Sub-second generation for 512x512 on commodity GPUs
Prompt-faithful for short, descriptive prompts (it's less good at 50-word abstract poetry, more useful for "cinematic shot of a fox in autumn light")

Practically, this means you can integrate it into a SaaS product without burning $0.04 per image like with proprietary APIs.

Why This Matters for Indie Builders

If you're shipping any product that needs image generation — content tools, e-commerce mockups, design helpers — the unit economics of AI tooling determine whether you can run a free tier at all.

A few quick numbers (rough, varies by hosting):

Model	Cost per image (approx)
DALL-E 3 (OpenAI)	$0.04
Midjourney	~$0.02 (subscription only)
Stable Diffusion XL self-hosted	$0.005
Nano Banana Pro	~$0.001 - $0.003

That last row is the difference between "feature locked behind a paywall" and "give every visitor 10 free generations a day."

Try It Yourself (No Sign-up)

You can test Nano Banana Pro for free at https://nanoai.run — there's a creation workbench that takes a text prompt and returns an image in a few seconds. Useful both as a sanity check on the model and as a reference for building your own integrations.

A couple of prompts that consistently produce good results in my experiments:

cinematic close-up of a fox in golden hour, photorealistic, 35mm film
flat illustration, retro 90s computer terminal on a wooden desk, warm light
studio product shot, ceramic coffee mug on marble surface, soft shadows

The "prompt that works" pattern I've found: short subject + lighting cue + medium/style anchor. Nano Banana Pro responds well to that recipe and tends to overcook on long, layered prompts.

What I'd Watch For

A few honest critiques after using it for a while:

Hands and text — still occasionally rough, like most diffusion models in this size class
Long prompts — the model is best at "short and concrete," not "describe the entire scene with mood, era, and 5 named subjects"
Style fidelity for specific artists — intentionally avoided in training (which is fine, but means you can't replicate a specific living artist's style)

None of these are dealbreakers. They're the predictable trade-offs of going from a giant model to a "nano" one.

Closing Thoughts

The interesting story in 2026 isn't that AI image generation got better. It's that it got cheaper and faster — and suddenly indie builders can ship features that were prohibitively expensive a year ago.

If you want to play with the model directly, the free workbench is at nanoai.run. I'd love to hear what prompts you find that work well — drop them in the comments.

Posted as an open exploration, not a product pitch — I just think this category is genuinely worth paying attention to.

DEV Community: Evan Li

Designing a quiz result that shows its limits

What is the primary output?

What does a comparison claim?

What happens when a result is shared?

Building an AI Image Generator SaaS in 2026: My Tech Stack and Lessons

The Stack

The Critical Decision: Where Do You Run the Model?

Option 1: Hosted API (OpenAI, Stability)

Option 2: Self-host on RunPod / Modal / Replicate

Option 3: Use a small efficient model via cheap inference provider

Free Tier Strategy

Don't offer "unlimited free"

Don't offer "10 free per month"

Do offer "10 free per day"

The Bot Abuse Problem

What I'd Skip in 2026

Lessons (the hard-won ones)

Try the Stack in Action

Closing

Nano Banana Pro vs DALL-E 3 vs Midjourney: A Practical Comparison From Someone Who Actually Uses All Three

TL;DR Up Front

What I Actually Use Each For

Quality Comparison: 5 Actual Prompts

1. cinematic close-up of a fox in autumn light, photorealistic

2. flat illustration, retro 90s computer terminal, editorial style

3. studio product shot, ceramic coffee mug on marble surface

4. cyberpunk courier, full body, neutral pose, concept art

5. Tokyo back alley at night, light rain, atmospheric

Score Card

Where Nano Banana Pro Surprises

What I'd Recommend

Closing

The Real Cost of Adding AI Image Generation to Your SaaS in 2026

The Setup

What Actually Drives the Difference

1. Model size

2. Hosting margin

3. Quality-cost tradeoff (the honest part)

What I'd Do (Practical Advice)

Try a Cheap Model Before You Commit

Closing Thought

5 Prompt Recipes That Reliably Produce Great Images on Nano Banana Pro

Why Recipes Beat Random Prompting

Recipe 1: The "Cinematic Portrait"

Recipe 2: The "Flat Editorial Illustration"

Recipe 3: The "Studio Product Shot"

Recipe 4: The "Mood Landscape"

Recipe 5: The "Character Card"

Pattern: Anchor Words

Try Them

How Nano Banana Pro Model Is Quietly Changing AI Image Generation in 2026

What is Nano Banana Pro?

Why This Matters for Indie Builders

Try It Yourself (No Sign-up)

What I'd Watch For

Closing Thoughts

1. `cinematic close-up of a fox in autumn light, photorealistic`

2. `flat illustration, retro 90s computer terminal, editorial style`

3. `studio product shot, ceramic coffee mug on marble surface`

4. `cyberpunk courier, full body, neutral pose, concept art`

5. `Tokyo back alley at night, light rain, atmospheric`