Why lm-arena Beats Fal.ai (And Standalone Sora)

Part of the Murder Mystery 1926 project

The Problem

I needed a video teaser for my murder mystery game.

Requirements:

• 1920s aesthetic
• Noir atmosphere
• Specific props (piano wire, bloodstains, murder weapon vibes)
• Professional enough to hype 10 people for Christmas dinner murder

My resources:

• Fal.ai budget (not unlimited, but enough to test)
• Access to multiple video generation tools
• 2 weeks timeline
• Zero video production experience

My expectations: Generate some clips, edit together, done.

Reality: Most tools gave me absolute garbage.

The Fal.ai Experience

What I Tried

Fal.ai has MANY video generation models:

• Minimax
• Kling
• Luma
• Runway
• Various Sora alternatives

What I did:

• Tested them systematically
• Wrote detailed prompts
• Tried different prompt styles
• Adjusted parameters
• Generated multiple variations
• Spent actual money

Results: 🗑️ Shit. Consistently shit.

Why It Sucked

Not Fal.ai’s fault specifically - the platform works fine.

The problem:

• Models didn’t follow prompts accurately
• 1920s aesthetic = random interpretation
• “Piano wire” = ???
• Noir lighting = sometimes just dark, sometimes neon???
• Parameter control felt limited
• No way to compare models side-by-side efficiently

Example prompt:

Extreme macro close-up, broken piano wire approximately 30cm length
coiled on dark wood surface (mahogany desk), wire diameter 0.8mm
visible in sharp detail, dark brownish-red stains on sections of wire
(dried blood implication)...

Result: Not even close. (See the blooper video in the project page.)

Then I Tried Standalone Sora

Why Sora: Heard it was the best for realistic, cinematic generation.

Access: Through official OpenAI interface.

Prompt: Same detailed, parameter-heavy prompt.

Result: Still… not great?

Better than some Fal.ai models, but:

• Didn’t nail the aesthetic consistently
• Still had weird interpretations
• Limited control over output
• Expensive per generation

Frustration level: High.

Then Someone Suggested lm-arena

What is lm-arena?

lm-arena.ai = Platform for comparing LLMs and multimodal models side-by-side.

But it has video generation.

With multiple models. Including… Sora.

My reaction: “Wait, isn’t that the same Sora I just tried?”

Narrator: It was. But the experience was completely different.

The lm-arena Experience

Why It’s Better

1. Side-by-side comparison

You generate with multiple models at once.

Same prompt → 2-4 different models → instant comparison.

This is HUGE because:

• You see which model understands your prompt better
• You learn which models work for your aesthetic
• You don’t waste time/money on single generations
• You can iterate faster

Example:

• Prompt about piano wire + bloodstains
• Generate with Sora, Kling, Runway simultaneously
• Pick the best result
• Adjust prompt based on what worked
• Repeat

This workflow is SO MUCH FASTER.

2. Better Sora results (somehow???)

I don’t know why. I really don’t.

Same model. Same prompts. Better outputs.

Theories:

• Different inference parameters?
• Different sampling settings?
• Platform optimization?
• Pure luck?
• The universe decided to help me?

I have no idea. But it worked.

3. The voting system teaches you

lm-arena has a voting system (which model’s output is better).

Why this matters:

• You start seeing patterns in what works
• You learn which models excel at what
• You develop prompt intuition faster
• Community votes = implicit feedback

It’s like having a focus group for your prompts.

4. Multiple accounts strategy

The hack: I had access to multiple Discord accounts (thanks friends!).

Why this helped:

• Parallel generation across accounts
• More attempts per hour
• Faster iteration cycle
• Test multiple prompt variations simultaneously

Ethical? Debatable. Effective? Extremely.

(Note: Check lm-arena’s terms of service. I’m just documenting what I did.)

The Results

What I Generated

Total clips generated: 20+ across all attempts

Usable clips: 5-6 high quality

Final teaser: 1 clip I fell in love with (and actually used!)

Bonus: 1 clip I loved but didn’t fit the teaser (saved for future use)

The Winning Workflow

1. Write detailed prompt

• Specific parameters (camera angle, lighting, movement)
• Reference aesthetics (noir, 1920s, cinematic)
• Technical details (macro lens, 8K, color grading)

2. Generate on lm-arena with multiple models

• Sora (usually the winner)
• Kling (sometimes surprising)
• Runway (hit or miss)

3. Pick the best result

• Vote on lm-arena
• Download the winner
• Note which model worked

4. Iterate on the prompt

• Adjust based on what worked/didn’t
• Test variations
• Refine parameters

5. Repeat until satisfied

Total time: ~4 hours of active work (spread over 2 days)

Total cost: Way less than Fal.ai burn

Direct Comparison

Fal.ai

✅ Pros:

• Many models in one place
• Simple interface
• Pay-per-generation pricing

❌ Cons:

• No side-by-side comparison
• Slower iteration (one model at a time)
• Higher cost for same results
• Hard to learn which model fits your needs

Standalone Sora

✅ Pros:

• Official OpenAI interface
• High quality model
• Reliable access

❌ Cons:

• Expensive
• Single model (no comparison)
• Slower learning curve
• Results weren’t as good as lm-arena (for me, somehow???)

lm-arena

✅ Pros:

• Side-by-side comparison (game changer)
• Multiple models simultaneously
• Learn faster which models work for you
• Sora results were better (I don’t know why!)
• Community voting = feedback loop
• More efficient use of time/budget

❌ Cons:

• Interface not as polished as dedicated platforms
• Requires understanding of multiple models
• Rate limits (unless you have multiple accounts 👀)
• Not designed primarily for video generation (but works!)

What I Learned

1. Platform matters as much as the model

Same model (Sora), different platforms, different results.

I don’t fully understand why, but it’s real.

Possible factors:

• Inference parameters
• Sampling settings
• Platform optimization
• Random seed differences
• API vs web interface differences

Lesson: Don’t give up on a model after one platform fails. Try different interfaces.

2. Side-by-side comparison accelerates learning

Before lm-arena: Generate → evaluate → adjust → generate → repeat

• Slow feedback loop
• Hard to know which model suits your style
• Expensive trial and error

With lm-arena: Generate (multiple) → compare → pick best → adjust → repeat

• Fast feedback loop
• Learn model strengths quickly
• Efficient experimentation

It’s like A/B testing for AI generation.

3. Prompt engineering is model-specific

What worked for Sora: Detailed, technical, parameter-heavy prompts

What worked for Kling: More conceptual, aesthetic-focused prompts

What worked for Runway: Shorter, action-focused prompts

You can’t learn this without comparison.

4. Multiple accounts = parallel experimentation

Ethically gray? Yes.

Practically useful? Extremely.

If you have friends willing to lend accounts (with their consent!), parallel generation speeds up iteration massively.

But: Check terms of service. Don’t abuse systems. Be respectful.

5. Budget on the wrong platform = wasted money

I spent money on Fal.ai with mediocre results.

Then got better results on lm-arena (which has free tier + voting credits).

Lesson: Test platforms before committing budget.

Practical Recommendations

If You’re Starting Video Generation

1. Start with lm-arena

• Free tier available
• Test multiple models
• Learn what works for your style
• Don’t commit budget until you know which models you need

2. Use side-by-side comparison aggressively

• Generate with 3-4 models at once
• Vote honestly (helps the community)
• Take notes on which models excel at what

3. Write detailed prompts

• Technical parameters (lens, lighting, camera movement)
• Aesthetic references (noir, 1920s, cinematic)
• Specific details (props, colors, textures)
• Negative prompts (what you DON’T want)

4. Iterate quickly

• Don’t expect perfection on first try
• Test variations
• Learn from failures
• Refine prompts based on results

If You’re Frustrated with Fal.ai/Standalone Tools

Try lm-arena.

Seriously.

Same models. Different experience.

Why it might work better:

• Side-by-side comparison changes workflow
• Voting system provides implicit feedback
• Community-driven model selection
• Potentially different inference settings

The Murder Mystery Teaser

Final result: 30-second teaser video

Tools used:

• lm-arena (Sora) for main footage
• Adobe Firefly for SFX
• Gemini for voice clips
• kdenlive for editing

Total cost: Fraction of what I burned on Fal.ai

Quality: Good enough to hype 10 people for a Christmas murder mystery

Watch it: Murder Mystery 1926 project page

Honest Disclaimer

I don’t know WHY lm-arena worked better.

Maybe:

• Inference settings
• Platform optimization
• Random luck
• Confirmation bias
• The universe conspiring to help my murder mystery game

All I know: Same model, different platform, better results (for me).

Your mileage may vary. Test for yourself.

Final Thoughts

Video generation tools are NOT plug-and-play.

You will:

• Generate garbage
• Waste money
• Get frustrated
• Question your prompts
• Question the tools
• Question your life choices

But:

• Some platforms work better than others
• Side-by-side comparison accelerates learning
• Prompt engineering improves with practice
• Eventually you get results you’re proud of

For me, lm-arena was the breakthrough.

Maybe it’ll be yours too. Maybe not. Only one way to find out.

Part of the Artifactum series - Murder mysteries built with AI assistance.

Next: The full video production pipeline (coming soon)

Tools mentioned:

• lm-arena.ai - Multi-model comparison platform
• Fal.ai - AI model hosting platform
• Sora - OpenAI’s video generation model
• Adobe Firefly - SFX generation
• kdenlive - Open-source video editor

Disclaimer: Not sponsored. Just sharing what worked (and didn’t) for my project.

Generated weird AI video clips? Show me! I want to see the bloopers. 🎬