Gen-4.5

I tested Runway's Gen4 Turbo across 30+ prompts spanning every category from physics simulations to human avatars. What I found was a model that gets you results fast but asks you to accept some rough edges in return.

After weeks of testing, Gen4 Turbo scores 50.47/100 overall. That's middle-of-the-road territory. The model delivers impressive speed and handles certain tasks beautifully, but it stumbles on the fundamentals too often to recommend without reservations. If you need fast previsualization or rough drafts, it works. If you're after polished final assets, keep looking.

The Physics Problem

Gen4 Turbo's physics simulation is where things get messy. Scoring just 47/100, it's clear the model doesn't quite understand how the world works.

Water looks impressive when Gen4 Turbo gets it right. Reflections work, fluid dynamics show some sophistication. But consistency? That's the problem. One moment you're seeing convincing buoyancy, the next you're watching fabric phase through solid objects like a video game glitch.

Humans Look Good (Mostly)

Scoring 56/100 for human avatars, Gen4 Turbo does better with people than I expected from a turbo model.

I watched a pianist play a slow arpeggio. The hands moved naturally, fingers depressed the correct keys in sequence, and the reflection on the keyboard looked polished. The actor's head nodded along with the music. It felt believable. The only letdown? Text fidelity on the piano itself rendered as gibberish, but that's a separate issue.

An anime barista pouring latte art showed the model's range. The arm proportions stayed correct throughout the pour, the wrist rotation looked precise, and by second 6, you could see a heart pattern forming in the cup. The smile remained consistent. For a stylized character in motion, this was solid work.

But here's where it gets weird. On a stage performance prompt, the actor looked excited and energetic. Great. Except her eyes turned demonic. At first, I thought it was just shading and mascara. But as the camera pushed closer, the whites of her eyes never resolved correctly. Combined with that AI-perfect smooth skin effect, the uncanny valley crept in hard.

Then there was the clapping issue. I explicitly asked for the main character to clap. She didn't. Her hands stayed frozen mid-gesture while background actors clapped rhythmically like they were all synced to the same metronome. It looked bizarre. The actor seemed to want to clap but physically couldn't. That's a prompt adherence failure that tanks the score.

Prompt Adherence Is Hit or Miss

At 50/100, Gen4 Turbo's prompt adherence sits right in the middle. Sometimes it nails complex instructions. Other times, it completely ignores them.

The red ceramic mug test was brutal. I asked for a teapot to pour tea into a red mug sitting next to a yellow teapot, with a green apple on a saucer between them. Simple enough, right? Gen4 Turbo poured the water onto the green apple instead. Multiple attempts, same failure. It's like the model couldn't parse the spatial relationship between "pour into the mug" and "the apple is just sitting there, don't touch it."

An anime baker scene asked for the character to lift a pink cake above her head, rotate it clockwise once, and return it to the counter by second 5. The arm pierced through the glass counter. The cake never went above her head. Partial execution at best.

When prompts stay simple, Gen4 Turbo performs much better. A metronome ticking beside three glass marbles with soft window light? It got close. The metronome moved (though not at a perfectly regular cadence like a real metronome would), and the marble count was wrong, but the reflections on those marbles looked fantastic. You could see light refraction, floor reflections, the whole deal. Materials and textures impressed me here, even if the logical details fell apart.

Scene Consistency: Acceptable, Not Great

Scoring 48/100, Gen4 Turbo handles temporal stability okay but not spectacularly.

A teal-braided anime student stirring a pink beaker showed decent consistency. The steam rose steadily, the braid lines stayed crisp without crawling across the frame. But pause at any moment and you'd see warping around the edges. The reflections on the beaker's surface shimmered more than they should. It's watchable but clearly AI-generated.

The model does better with simpler scenes. A brass metronome scene had solid lighting that stayed constant throughout. The metronome itself moved with believable physics (spring-back motion looked right), even if the timing wasn't perfect. For scene consistency, this scored 65, which is respectable.

Where it breaks down is complex motion. Fabric patterns tear and shift. Small details dissolve behind moving hands. You get that telltale AI shimmer on fine textures. None of this makes the video unwatchable, but it does make it obviously synthetic.

Logic Consistency Falls Apart

At 43/100, this is Gen4 Turbo's weakest dimension.

I asked it to show someone stacking three labeled books and returning them to their original position. What happened? The actor picked up a book that split into two pieces mid-grab. He deposited part of it behind him on the table (where it shouldn't go), and the letter "A" appeared in multiple places at once. The spine texture looked amazing. The lighting was gorgeous. But the logical sequence made zero sense.

A fashion animation with two women having tea started strong. Shadows pointed in the right direction, the composition looked natural. Then as one woman set down her cup, her hand transformed into a saucer. Both characters suddenly held cups in impossible configurations. One woman had a pink trim on her sleeve. Not the other sleeve, just one. These aren't physics errors, they're logic errors. The model hallucinates details that break the internal consistency of its own scene.

An Asian gentleman on a white cyc scored higher (middle of the road) because at least the logic held. Nothing weird happened. The person mumbled, the lights made sense, and the cyc stayed white. Boring? Yes. Logically consistent? Also yes.

Text Fidelity: A Known Weakness

Gen4 Turbo scored 50/100 here, which matches what most video models achieve. Text is hard.

Up close, the model can pull off readable text. "Sunrise Cafe" appeared in two lines when I expected one, but you could read it. From a distance or in reflective environments? Complete breakdown. The second instance of "Cafe" turned into incomprehensible symbols. A reflective window test produced gibberish that clearly said "AI generated this."

One bright spot: A "LuminArchive" effect scored 75. I hated the aesthetic (way too flashy and artificial), but the letters were legible. Large, head-on text works for Gen4 Turbo. Anything else is a gamble.

The CG Look Is There (Sometimes)

Feature CG style scored 50/100. Gen4 Turbo can produce that Pixar-adjacent look, but it's not consistent.

A living room kitchen scene had excellent wood textures and metallic surfaces. The objects looked photorealistic in a stylized way. That's what you want from feature CG. But the scene logic made no sense. Pots everywhere, half industrial kitchen, half living room. A convection oven appeared next to what might be a crock pot. The model prioritized visual appeal over spatial coherence.

A photorealistic animal render with a furry creature showed strong subsurface scattering (well, the turbo version of it). The fur looked decent. But the creature barely moved, giving it that stiff puppet quality. For a polished Pixar-level result, you need motion that matches the visual quality. Gen4 Turbo couldn't deliver both.

Animation: A Hidden Strength

This was a pleasant surprise. At 55/100, Gen4 Turbo handles Japanese anime better than expected.

A wind-blown fabric scene with a character's hair and clothing reacting to wind looked like it came from a high-budget animation. The only issue was massive artifacting where the fabric met the wind effect. If that had been cleaner, this would have scored in the 70s easily.

Two women having tea in anime style showed natural shadows, proper composition, and believable character interaction. The teacups looked right, the setting worked. The elbow had some issues, and one sleeve had pink trim while the other didn't, but the overall animation quality was strong.

Gen4 Turbo seems to have been trained heavily on anime content. When you give it that style, it delivers.

Animals Are Rough

Gen4 Turbo scored 50/100 on animals, and frankly, that feels generous.

A red panda prompt asked for the animal to sniff the air, blink twice, and gently swish its tail. Instead, the panda walked off. Okay, that's a prompt adherence issue. But it looked like a red panda, so partial credit. Then I noticed the tail. At three seconds in, the tail split into two tails. Then disappeared. Completely. That's not a minor glitch, that's a fundamental breakdown in understanding what a tail is.

An otter scene was worse. Multiple otters with too many fingers, water spitting from mouths in unnatural ways, and faces that looked more alien than otter. The behavior was okay, but the anatomy was wrong. For a model trying to simulate the world, getting basic animal anatomy right should be table stakes.

UX and Cinematography: Solid When It Works

Scoring 55/100, Gen4 Turbo's user experience is functional.

A camera motion test asking for a specific dolly move executed it correctly. The image looked good, the framing made sense, and the model did what I asked. For UX, that's a win. A fine dining dish with camera movement also worked well. The dish itself was nonsensical if you looked closely, but the camera work supported the action.

Where it falls apart is when complex logic meets camera movement. Background elements drift, reflections don't stay static when they should, and framing occasionally obscures the main subject when it shouldn't.

What This Means for You

Gen4 Turbo is a tool for speed, not polish.

If you're doing rapid iteration, prototyping ideas, or creating animatics, the 30-second generation time for a 10-second clip is hard to beat. The model gives you enough fidelity to make decisions. You can see if a scene works, if the composition feels right, if the motion reads clearly.

If you're delivering final assets, Gen4 Turbo isn't ready. The logic failures alone disqualify it for professional use. A hand turning into a saucer, a tail splitting in two, a teapot pouring onto the wrong object—these aren't minor artifacts. They're fundamental failures that break viewer immersion.

The model shines in specific niches. Japanese anime? Surprisingly good. Simple human avatars in controlled environments? Decent. Basic camera movements? Reliable. But ask it to handle complex physics, maintain logical consistency across a scene, or execute precise spatial instructions, and it struggles.

Final Thoughts

I wanted to love Gen4 Turbo. The speed is genuinely impressive, and when it works, it works well. But testing it across 30+ prompts revealed a model that's still figuring out the basics.

Runway clearly optimized for speed over accuracy. That's a valid trade-off for certain use cases. But it means you're constantly making compromises. Can you live with physics that breaks under scrutiny? Can you accept that your character's anatomy might dissolve mid-scene? Can you work around a model that sometimes just ignores your instructions?

For previsualization and rough drafts, yes. For anything else, probably not.

---

Score Summary

• Overall Score: 50.47/100
• Prompt Adherence: 50/100 (Good - Accurate overall with brief failures)
• Scene Consistency: 48/100 (Acceptable - Moderate flickering in fine details)
• Physics: 47/100 (Acceptable - Moderate inaccuracies during heavy motion)
• Human Avatar: 56/100 (Good - Anatomy generally correct with brief glitches)
• UX: 55/100 (Good - Solid usability with minor pacing issues)
• Animation: 55/100 (Good - Solid timing with minor jitter)
• Live Action: 55/100 (Good - Realistic overall with minor uncanny moments)
• Objects: 48/100 (Acceptable - Moderate errors, usable for draft review)
• Logic Consistency: 43/100 (Below Average - Noticeable errors but intent recognizable)
• Animals: 50/100 (Good - Solid anatomy with occasional slips)
• Text Fidelity: 50/100 (Good - Clear typography with brief shimmer)
• Actor Performance: 50/100 (Good - Believable acting with minor stiffness)
• Feature CG Style: 50/100 (Good - Pleasant style with minor lighting mismatches)

Verdict: Gen4 Turbo delivers speed at the cost of reliability. Use it for rapid iteration and previsualization, but don't expect final-quality assets without significant cleanup.