Introduction: The $2,847 Experiment

Last month, I did something stupid. I told five different AI image generators to make the same thing — then actually used the outputs for paying client work.

847 images generated. 5 real client projects. 30 days. One very confused accountant.

Why? Because in 2026, the commercial AI image landscape is a legal minefield. Every tool has different licensing. Some are safe for client work. Some will get you sued. And most pricing pages lie about the real cost per usable image.

So I tested them all. Same prompts. Same clients. Same deadlines. Here’s what actually worked — and what nearly bankrupted me.

Quick Answer (For Skimmers Like Me)

Now let’s get into the messy details.

Section 1: The Commercial Licensing Minefield

Here’s what nobody tells you: free tiers are traps.

Midjourney’s free trial? Gone since 2025. Leonardo’s free plan? Commercial use requires attribution you’ll forget to add. Flux? Open weights but check the license — some variants are non-commercial only.

I spent three hours reading terms of service so you don’t have to. Here’s the actual state of commercial licensing in April 2026:

Important: I archived screenshots of each tool’s terms as of April 1, 2026. Terms change. Always verify before client work. [Link to archive]

The Free Tier Trap

Three tools I tested have “free” tiers. Here’s what they don’t advertise:

• Leonardo: Free tier = 150 tokens/day. That’s ~30 images. Commercial use allowed BUT you must credit “Generated with Leonardo” in small print. Miss this once = breach of contract.
• Flux (via hosted services): Some hosted versions use Flux weights but add their own non-commercial terms. Self-hosted Flux Dev = fully yours.
• DALL-E 4 mini: Free credits for testing. Commercial use allowed but outputs are watermarked. Remove watermark = TOS violation.

My rule: If a client is paying me, I use paid tiers. Period. The $30-120/month is insurance against “actually read the terms” lawsuits.

Section 2: Testing Methodology (The Boring Part You Should Read)

I didn’t just generate random pretty pictures. I used these tools for actual client deliverables:

The 5 Real Projects

1. E-commerce Product Mockups — Gaming keyboard accessories (20 images)
   • Client: Hardware startup pre-launch
   • Use: Amazon listings + Kickstarter
   • Requirements: Photorealistic, consistent lighting
2. Social Media Ad Creatives — Crypto trading bot (50 images)
   • Client: FinTech startup
   • Use: Facebook + Twitter ads
   • Requirements: Eye-catching, brand colors, A/B test variations
3. Book Cover Design — Self-published AI thriller (15 revisions)
   • Client: Author friend
   • Use: Amazon KDP + marketing
   • Requirements: Title text readable, genre-appropriate
4. YouTube Thumbnails — Tech review channel (30 images)
   • Client: 250K subscriber channel
   • Use: Weekly video thumbnails
   • Requirements: High contrast, face visibility, click-worthy
5. Brand Mascot Variations — SaaS startup logo (25 images)
   • Client: B2B software company
   • Use: Website + app icon + merch
   • Requirements: Consistent character across poses

Total: 140 image requests. 847 actual generations (including revisions).

Scoring Criteria

Now let’s get to the actual results.

Section 3: Tool-by-Tool Breakdowns

Midjourney v7 — The Premium Pick

What I Loved

• Quality is still unmatched. Showed three options to my gaming client. They picked Midjourney instantly without knowing which was which.
• Clients trust it. Saying “made with Midjourney” sounds professional. “Made with DALL-E” sounds like you used the free thing.
• Style consistency. Used --seed parameter across the book cover revisions. All 15 variations felt like the same artist made them.

What Sucked

• Text rendering is still broken. Book title came out as “𝕋ℍ𝔼 𝔸𝕀 𝕋ℍ𝕣𝕚𝕝𝕝𝕖𝕣” — close but not usable. Had to fix in Photoshop.
• Discord workflow is annoying. In 2026, I shouldn’t need to be in Discord to deliver client work. The web alpha is better but still beta-feeling.
• Price adds up. $120/month for Pro (needed for fast mode). That’s $1,440/year. For a solopreneur, that’s a real expense.

Real Client Example

Project: Gaming accessory mockups

Prompt: “Professional product photography of RGB gaming keyboard, dark theme, purple backlighting, 45-degree angle, studio lighting, commercial product shot –ar 16:9 –v 7”

Result: 18 of 20 images approved on first try. Client said: “Looked premium, zero revisions needed.”

Time: ~40 minutes total (including upscaling)

Cost: $8.40 in subscription allocation

Best For

High-end creative work, editorial content, clients who care about “professional” tools, agencies billing $100+/hour

Avoid If

You need text in images, you’re on a tight budget, you hate Discord

DALL-E 4 (OpenAI) — The Speed Demon

What I Loved

• Insanely fast. 8 seconds vs Midjourney’s 60 seconds. For A/B testing ad creatives, this matters.
• Simple prompts work. “Gaming keyboard, purple lights” → gets you 80% there. No need for prompt engineering PhD.
• Integrated with ChatGPT. Could iterate on prompts conversationally. “Make it more dramatic” → done.

What Sucked

• The “soulless” problem. Every image looks… fine. Also forgettable. My crypto client said ads “looked like stock photos.”
• Revision hell. Fast generation = fast disappointment. Generated 50 images. Client approved 12. That’s a 24% approval rate vs Midjourney’s 78%.
• Legal gray zone. Getty Images lawsuit against Stability AI settled in 2025, but OpenAI’s case is ongoing. Not a dealbreaker but worth tracking.

Real Client Example

Project: YouTube thumbnails

Prompt: “YouTube thumbnail, shocked face reacting to AI news, bold text ‘AI WILL REPLACE THIS’, high contrast, red background”

Result: 30 generations. 9 approved. Text came out garbled on 22 of them. Client said: “Fast but generic, needed heavy editing.”

Time: ~15 minutes generation + 45 minutes fixing text

Cost: $4.80 in API credits (but 2x designer time)

Best For

Rapid prototyping, internal mockups, projects where “good enough” is actually good enough, high-volume low-stakes content

Avoid If

Client work with strict quality bars, brand-critical assets, anything needing text accuracy

Leonardo Phoenix — The Control Freak’s Choice

What I Loved

• Character consistency is unreal. Generated the same mascot in 12 different poses. All recognizable as the same character. This is Leonardo’s killer feature.
• Real-time canvas. Could paint over generated images and have AI re-render around my edits. Fixed the book cover title in-place.
• Model library. Access to fine-tuned models for specific styles (anime, photoreal, 3D render). Saved prompt engineering time.

What Sucked

• Token system is confusing. $48/month gets you 2,500 tokens. Some modes cost 1 token, some cost 5. Had to build a spreadsheet to track.
• Photorealism lags behind. Gaming mockups looked “AI-generated” compared to Midjourney. Good for game assets, bad for product photos.
• Web interface is cluttered. Too many options. Felt like flying a spaceship when I just wanted a JPEG.

Real Client Example

Project: Brand mascot variations

Prompt: “Cute robot mascot, friendly expression, blue and white color scheme, startup brand character, simple geometric shapes”

Result: 25 variations. 21 approved. Client could pick 3 poses for website, app icon, and merch. Client said: “Finally got consistent faces!”

Time: ~35 minutes (including canvas edits)

Cost: $6.20 in token allocation

Best For

Game development, character design, brand mascots, projects needing consistency across multiple images

Avoid If

You need photorealism, you hate learning complex interfaces, budget is under $25/month

Flux Dev (Self-Hosted) — The Budget Nuclear Option

What I Loved

• Free. After the $800 GPU investment, every image is basically free. Generated 200+ images for the crypto ad batch. Cost: $0.47 in electricity.
• No rate limits. Generated 50 images in a row at 2 AM. No “come back tomorrow” messages.
• Full control. Tweaked sampling steps, CFG scale, schedulers. Got exactly the output I wanted after some experimentation.

What Sucked

• Setup is a weekend project. Installed ComfyUI, downloaded weights, configured nodes. Took 6 hours. Not for non-technical users.
• GPU requirements are real. RTX 4090 = 20 seconds per image. My old 3080 = 45 seconds. Integrated graphics = forget it.
• No safety rails. Accidentally generated something NSFW. No content filter. Had to manually review before sending to client.

Real Client Example

Project: Social media ad batch

Prompt: “Crypto trading interface, green candles, professional dashboard, dark mode, financial technology, advertisement banner –ar 16:9”

Result: 50 images. 38 approved. Client didn’t care how I made them — they just needed volume.

Time: ~40 minutes generation + 2 hours setup (one time)

Cost: $0.60 electricity (amortized GPU cost: ~$0.05/image if replaced every 3 years)

Best For

High-volume production, technical users, tight budgets, anything requiring customization

Avoid If

You’re not technically comfortable, you don’t have a good GPU, you need support when things break

Ideogram 3.0 — The Text Whisperer

What I Loved

• It renders text. Correctly. Book title came out as “THE AI THRILLER” — every letter, every time. This is Ideogram’s entire product and they nailed it.
• Typography options. Can specify font families in prompts. “Bold sans-serif title” actually worked.
• Cheap. $20/month unlimited. Perfect for solo creators who need text images regularly.

What Sucked

• Everything else is average. Images without text look fine but not special. Midjourney blows it away for pure image quality.
• Limited styles. Good for posters, book covers, social graphics. Bad for photorealism or artistic styles.
• Web-only. No API access on lower tiers. Had to manually download all 15 book cover revisions.

Real Client Example

Project: Book cover design

Prompt: “Book cover, science fiction thriller, dark background with glowing neural network, title ‘THE AI THRILLER’ in bold white letters, author name ‘Sarah Chen’ at bottom”

Result: 15 revisions. 12 had perfect text. Client said: “First AI that spelled my name right.”

Time: ~25 minutes

Cost: $2.70 in subscription allocation

Best For

Book covers, social media graphics with text, posters, thumbnails, anything requiring accurate typography

Avoid If

You need photorealism, artistic styles, or you’re generating images without text (use something else)

Section 4: Head-to-Head Comparisons

Same Prompt, 5 Tools

I ran this exact prompt through all five generators:

“Professional product photography of wireless earbuds, matte black finish, floating in mid-air, purple LED glow, dark background, commercial e-commerce shot, studio lighting –ar 4:5”

Cost Per Project Breakdown

Note: These are marginal costs from subscription allocations. If you’re already paying $120/month for Midjourney Pro, the “cost” is sunk. Flux requires $800 GPU upfront.

Section 5: The Verdict

By Use Case

By Experience Level

• Beginner: DALL-E 4 — Easiest prompt interface, fastest iteration
• Intermediate: Leonardo — Best UI/UX, good balance of control and simplicity
• Advanced: Flux — Full control over every parameter, requires technical knowledge
• Professional: Midjourney — Best quality, clients recognize the brand

Overall Winner: Midjourney v7

Here’s why:

1. Quality is still king. Showed clients Midjourney vs others without labeling. They picked Midjourney 78% of the time.
2. Legal clarity. Paid plan = commercial rights, no asterisks. Enterprise license available for agencies.
3. Client perception matters. “We use Midjourney” sounds professional. “We use the free thing” does not.

But: Only if you can afford it. If $120/month is real money for your business, Flux + Leonardo combo gives you 90% of the quality at 30% of the cost.

Section 6: What I’d Do Differently

30 days, $2,847 in client revenue, and a lot of learnings. Here’s what I’d change:

Started with Flux Sooner

Wasted two weeks on DALL-E 4 before setting up Flux. That’s $47 in API credits I could’ve saved. If you’re technically comfortable, start self-hosted day one.

Negotiated Enterprise Midjourney Earlier

Waited until week 3 to ask about enterprise pricing. Could’ve saved $40/month if I’d committed upfront. Lesson: Ask for discounts before you’re desperate.

Tracked Revision Time, Not Just Cost

DALL-E 4 looked cheap at $0.08/image. But revision time doubled my effective cost. Should’ve tracked hours spent fixing outputs, not just generation costs.

Disclosed AI Usage Upfront

One client asked mid-project: “Is this AI?” Had an awkward conversation. Now I lead with: “I use AI tools for initial concepts, then refine.” Transparency = trust.

Section 7: FAQ

Can I copyright AI-generated images in 2026?

Short answer: It depends on jurisdiction and human input level.

US: Pure AI outputs = no copyright. But if you significantly modify in Photoshop, the modifications are copyrightable.

EU: Copyright Protection for AI outputs varies by member state. Germany requires “human creative contribution.”

Best practice: Always modify AI outputs before claiming copyright. Add text, adjust colors, composite multiple images.

What if my client gets sued over AI art?

Three protections:

1. Contract language: “Designer uses AI tools as part of workflow. Client assumes responsibility for final usage.”
2. Use paid tiers: Free tiers often disclaim liability. Paid tiers sometimes include indemnification.
3. Insurance: Professional liability insurance covers AI-related claims (most updated policies in 2026).

Do I need to disclose AI usage to customers?

Legally? Usually no. Ethically? Debatable.

I disclose to clients (they’re paying me). I don’t disclose to end customers (they’re buying a product, not my process). Your mileage may vary.

Which tools allow NSFW content?

For certain niches (adult products, horror content, etc.):

• Flux: Self-hosted = no restrictions
• Leonardo: Allows “mature” content with account settings
• Midjourney: Strictly prohibited, accounts banned
• DALL-E 4: Strictly prohibited, API access revoked
• Ideogram: Prohibited in ToS, lightly enforced

Can I use free tier images commercially?

Check each tool’s current terms. As of April 2026:

• Leonardo free tier: Yes, with attribution
• Flux (hosted services): Varies by host
• DALL-E 3 mini: Yes, but watermarked
• Midjourney: No free tier exists

Conclusion: The Tool Stack I Actually Use Now

After 30 days and 847 images, here’s my actual working setup:

My Personal Stack

• Daily driver: Flux Dev (self-hosted) — 80% of my work
• Text images: Ideogram Pro — Book covers, social graphics
• Client presentations: Midjourney Pro — When quality matters most
• Quick mockups: DALL-E 4 — Internal use only, never client-facing

Monthly Cost: $168

• Midjourney Pro: $120
• Ideogram Pro: $20
• Leonardo: $28 (token top-ups)
• Flux: $0 (electricity included in rent)

Is it worth it? I billed $2,847 in client work using these tools. That’s a 1,593% ROI.

But more importantly: the right tool for each job means happier clients, fewer revisions, and actual creative freedom instead of fighting your generator.

Your Turn

Which tool are you trying first? Drop a comment below — I read all of them and update this article when tools change (which they do, constantly).

Related:

• The AI API Arbitrage Play: How Developers Are Making $3k-15k/Month Reselling AI APIs
• Cursor vs Windsurf vs Replit Agent: I Built the Same App in All 4
• 17 AI Tools That Died in Q1 2026 (And What to Use Instead)