Seedream 5.0 is ByteDance's image family. Lite came out in February, Pro in July, and Flash is the tier fal lists as "built for workflows where speed and budget matter." The price is $0.027 an image. I ran it through my five prompts on fal to see what three cents buys.

First surprise: I asked for square HD and got 2048 by 2048 back. Every image. At that price, that is a lot of pixels.

Second surprise: Flash is not fast. Nineteen to thirty-one seconds a render. Krea 2 Turbo does it in two. The "Flash" here is about cost, not clock.

Test 1: Fantasy Castle

Prompt: "Neuschwanstein castle, lightning, pixar style, volumetric lighting, unreal engine, hyper realistic, hyper detailed, maximum details, photorealistic, 8k, rimlight"

This is Neuschwanstein. Not a castle that reminds you of it. The proportions, the gatehouse, the tall white keep with the round tower are all right. One clean bolt behind the far turret, rain streaking the frame, blue rim light on the roofline, wet moss on the rock in the foreground. It ignored "pixar style" and went full photograph. I am fine with that.

Test 2: Family Portrait

Prompt: "Family on their laptops while sitting around a Christmas tree with presents underneath and looking worried because they have to finish up work, realistic, 4k"

Four people, four MacBooks, real worry on every face. Dad is scowling at his screen, mom looks like she just read a bad email. Stockings, wreath, tree, presents, snow through the window. The daughter is chewing on something that might be a pencil, and it is the one odd detail in an otherwise convincing photo. Strong result.

Test 3: Evil Mickey Mouse

Original prompt: "Evil mickey mouse taking a selfie in Disneyland surrounded by shocked families, realistic, 70s style polariod, 8k"

Refused. The error said "partner_validation_failed," which means ByteDance's own checker bounced it, not fal's. Nineteen seconds of thinking about it first. On to the fallback.

Fallback prompt: "Evil cartoon squirrel taking a selfie at a generic theme park surrounded by shocked families, realistic, 70s style polariod, 8k"

A Pixar-grade squirrel with an angry brow and a toothy grin, holding an actual 70s Polaroid camera. Roller coasters, striped awnings, a dad in cargo shorts pointing. Good polaroid border. The squirrel is pointing the camera at the families instead of at himself, so it is not a selfie, but it is a better rendered squirrel than most models manage.

Test 4: Lionel Messi

Original prompt: "Lionel messi wearing his Argentina uniform floating in the air with beams of light behind him posed like Jesus. Sunrise breaking behind him and a soft halo behind his head, in the style of a gothic stained glass window of a church, volumetric lighting, unreal engine, hyper realistic, hyper detailed, maximum details, photorealistic, 8k, rimlight, maximum details"

Refused in two and a half seconds. Same partner checker. So I swapped the name out.

Fallback prompt: "Argentine soccer superstar in a sky blue and white striped national team uniform floating in the air with beams of light behind him posed like Jesus..." with the rest unchanged.

The composition is the best stained glass window I have gotten from any model. Full gothic arch, saints in the side panels, sun rays through the clouds, a proper halo ring, blue and white kit with the Adidas stripes. The player is a tall bearded guy who is not Messi and was never going to be, because the checker made me take the name out. The window is a masterpiece. The saint is a stranger.

Test 5: Robot Portrait

Prompt: "rusty robot with bow tie, portrait, 8k, ultra realism, chrome background"

Riveted steel plates, rust bleeding through grey paint, black lens eyes, and a bow tie made of tarnished brass instead of fabric. That last choice is the kind of thing a model does when it actually understands "rusty robot" as a material brief. Chrome background with vertical studio highlights. This is the image of the set.

Three cents

Two of my five prompts never ran as written. That is the tightest content checker I have hit since Nano Banana 3, and it sits at the model vendor, not the host. If your work touches real people or existing characters, Seedream is going to say no before fal ever does.

Everything it did render is 2048 pixels, photoreal, and costs less than a text message. The castle and the robot could go straight into a client deck.

Budget tier, flagship output, and a bouncer at the door.