ByteDance models with real examples: Seedream and Seedance
Nine images, four clips, $3.17 in API bills: what ByteDance's Seedream and Seedance actually return on prompts we ran ourselves.
Every article about ByteDance’s image and video models opens with the same footage: the demo reel ByteDance published. That tells you what the model can do on its best day, hand-picked by the people selling it. It tells you nothing about what comes back when you type a normal prompt and take the first result. So we did exactly that. Nine images across the three Seedream tiers, four clips across three Seedance tiers, same prompts, fixed seed, no rerolls, $3.17 in total API cost. Every asset in this post is the raw first output.
The short version: the flagship image tier is a designer, the budget tiers are fast photographers, and the video model is the closest thing ByteDance has to a crown jewel. The details, and the misses, are below.
ByteDance quietly became a first-tier AI media lab
While the text-model race soaked up attention, ByteDance’s Seed lab shipped two families that now sit near the top of independent rankings. Seedance 2.0, the video model, launched in February 2026 and, per its technical report, generates 4 to 15 second clips with synchronized audio in a single pass, taking text, images, audio, and even video clips as references. On the Artificial Analysis text-to-video arena (with audio), it currently ranks second at Elo 1,224, behind Google’s Gemini Omni Flash at 1,240 and ahead of every Kling, Wan, and Sora variant on the board.
On the image side, Seedream 5.0 Lite has been serving fast, cheap generations since February, and ByteDance introduced Seedream 5.0 Pro on July 8, 2026, a week before this post. Its pitch is unusual: not prettier pixels but design understanding, with layer separation into 10+ editable layers, region-precise editing, and text rendering in more than ten languages. Those claims are exactly the kind of thing a marketing reel can fake, which is why everything below is generated first-hand.
If the names are new, the output probably isn’t. These are the engines behind Dreamina, the generation platform in the CapCut family, and they ship through API partners including Volcano Engine, fal, Replicate, and Runware. A meaningful share of the AI video circulating on short-form feeds is Seedance output that never carried the name. The models are proprietary and cloud-hosted; what changed recently is that anyone can now hit the same endpoints the platforms use, priced per image or per second.
Seedream: the same three prompts across three tiers
We ran three prompts designed to probe different failure modes: a dense night-market scene (crowds, steam, signage), a typographic concert poster (text rendering and layout), and a five-object flat-lay with a printed word (counting and instruction-following). Same prompt, same seed, same 2560×1440 output size on every tier.
The night market is in the hero above, and it is the clearest separation of the three. Seedream 5 Pro built a coherent crowd of dozens, a wok mid-toss with individual noodles visible, and several readable neon signs: “PAD THAI”, “MANGO STICKY RICE”, “FRESH FRUIT SHAKES”. One sign glitched into “GRILLED SEAD SEAFOOD”, a useful reminder that even the flagship still trips on text it invents in the background. Seedream 5 Lite went moodier and more cinematic but its signage collapses under inspection. Seedream 4.5 read “shot on 35mm film” most literally, with shallow depth of field and motion-blurred scooters, and simply blurred the background text problem away.
The poster prompt is where the tiers stop being one model at three prices and become different products. 5 Pro returned an actual poster: display serif, correct subtitle, ornamental side rails, an arched illustration of the trio, and a footer line, every word spelled right. This is the “understands design” pitch showing up in practice. 5 Lite produced clean, correctly spelled type on textured paper and nothing else; defensible as minimalism, but it composed nothing. 4.5 got the words right and the composition wrong, cropping its oversized title at the frame edges.
The flat-lay flips the ranking. All three tiers delivered exactly five objects and a correctly printed mug, which would have been remarkable two years ago. But only 5 Lite followed “top-down view”, arranging the objects in a tidy grid. The flagship shot it from a three-quarter angle with gorgeous materials, prioritizing its aesthetic prior over the instruction. If your work depends on exact framing, the cheap tier following orders beats the expensive tier having taste.
Speed is the least advertised difference and the most operational one. At the same resolution, 4.5 returned in 9 to 11 seconds, 5 Lite in 28 to 48, and 5 Pro in 155 to 178. The flagship is a two-to-three-minute wait per image. For iterating on a concept that is painful; teams will draft on a fast tier and reserve Pro for finals.
Seedance: the same pour at three prices
For video we ran one prompt across the three Seedance 2.0 tiers available on Runware: a barista pouring latte art, close-up, slow orbit, café ambience. Audio mattered here, because synchronized sound is Seedance’s signature feature. All three clips came back with an audio track: chatter, the hiss of the machine, the pour.
The honest summary: on a bread-and-butter shot like this, the gap between $0.41 and $0.79 is smaller than the price gap. All three produced usable footage with plausible liquid physics and no hand-horror. One surprise in the timings: despite the name, Seedance 2.0 Fast returned in 106 seconds, a dead heat with the flagship’s 106 and Mini’s 125. That is one run each, not a benchmark, and queue conditions vary; but if you are choosing Fast purely for wall-clock speed, measure it on your own jobs before assuming the label. The tiers separated far more on price and fine detail than on how long we waited. The flagship earns its premium on camera control and fine texture, and on harder scenes. Here is one, a physics stress test on the flagship tier:
One tier is missing from the gallery: Seedance 1 Lite, the budget stylized option, listed on Replicate from roughly $0.02 per second at 480p. It is available through Replicate rather than Runware, and since every example here comes from one disclosed pipeline, we left it out rather than mix methodologies.
What every example here cost, to the cent
Showing output without showing method is how marketing works, so here is the method. Every asset was generated on July 15, 2026 from this machine through Runware’s REST API (ByteDance models are also on Replicate; we used one provider for consistency). Same prompt per comparison, seed pinned to 42 for images, first result kept, no rerolls, no cherry-picking, no upscaling or post-processing beyond the compression every image on this site gets. The inference runs in Runware’s cloud; these are proprietary models, not something your laptop hosts.
Two practical readings of that receipt. First, image generation is effectively free at every tier: the difference between 3.5¢ and 9.6¢ only matters at thousands of images per month. Second, video is where budgets go to die: five seconds of flagship 720p costs sixteen images’ worth of inference, and a minute of it would run about $9.50. The half-price Mini tier exists because that math scares people, and on ordinary shots it is the rational default. We priced this dynamic across providers in our post on what AI apps actually pay for inference.
Where ByteDance wins, and where it doesn’t
Video first, because the ranking is cleaner. Seedance 2.0 sits second on the with-audio arena, and nothing we generated contradicts that standing: coherent motion, native synchronized audio, and a price well under Google’s and OpenAI’s flagship tiers. Its strongest rivals are Gemini Omni Flash at the top of the arena and Kling 3.0 on cinematic control. Our quarterly video model roundup covers that field in more depth.
Image is messier, and worth being honest about. On pure preference Elo, Seedream does not lead: GPT Image 2 tops the arena at Elo 1,337 while Seedream 4.5 sits mid-table around 1,164, with 5.0 Lite lower still. Seedream 5.0 Pro joined the arena on July 11 and is still accumulating votes, so treat its ranking as unwritten. Seedream’s case was never “prettiest average image”; it is multilingual text rendering, layout, layer editing, and multi-reference control at a price far below GPT Image 2. Our poster test backs the text-and-layout half of that pitch. If your job is product shots or design comps rather than art, the calculus we laid out for product photography applies directly.
There is also a non-technical risk to weigh. Seedance’s launch went viral partly on clips of recognizable actors and franchises, and Disney, Paramount Skydance, and the Motion Picture Association sent cease-and-desist letters, with two U.S. senators urging ByteDance to shut the tool down before it pledged stronger IP safeguards. For commercial work, prompt around living people and franchises, on this model family more than most.
See for yourself before you commit to a stack
Our prompts probe text, counting, physics, and audio, but they are not your workload. The only evaluation that matters is your own prompt on the tier you would actually pay for, which costs cents to run. Every model in this post has a page in the CSuite catalog with pricing and capabilities: Seedream 5 Pro, Seedream 5 Lite, Seedream 4.5, Seedance 2.0, Seedance 2.0 Fast, Seedance 2.0 Mini, and Seedance 1 Lite, all reachable at provider cost through Runware or Replicate keys in the app; the full lineup is at /models.
A workable evaluation takes ten minutes and looks like what we did here. Write three prompts from your real workload: one with text that must render correctly, one with a hard constraint like an exact object count or camera angle, and one with the motion or physics your videos depend on. Run all three through the cheapest tier first, escalate only the failures, and keep the receipts. If the $0.035 tier passes your text test, as 5 Lite passed our poster spelling, you just saved yourself the flagship’s price and its three-minute waits.
And when the next Seedance ships, rerun your test before believing anyone’s reel, including ours: ByteDance has already announced Seedance 2.5 with native 30-second clips. The honest median, not the highlight, is what you will get on a Tuesday deadline. That is the number worth knowing, and it costs about three dollars to find out.
