The assistant asks what you left out and explains why each answer changes the result. Then it renders through 0G Compute, over a route you can verify on-chain: stills on Z-Image-Turbo, motion on MiniMax H3 or Seedance 2.5.
The assistant
DeepSeek V4 Flash runs a short interview instead of taking a prompt. One question at a time, always with the reason it matters, always with real options you can ignore. Five answers later you have a brief worth rendering, whether that is one still or a whole video.
A matte ceramic mug for a product page. Four angles.
What is it sitting on, and against? Backgrounds are where generated images fall apart. Naming the surface and the backdrop removes most of the mush. Answer: Worn wood.
How should it feel? Mood ties the grade, the palette, and the styling together so a set of images reads as one set. Answer: Warm.
Brief so far
Matte ceramic mug, centred with room to breathe, soft morning window light from the left, long falloff across the body, visible glaze texture, 4:5. On worn oak against a neutral backdrop. Warm and unhurried.
Generating a still
Subject, framing, light, surface, mood. The answers compose into a single prompt, and the whole set renders from that one prompt instead of five unrelated ones.
Generating a video
Subject, hook, setting, tone, payoff. The answers become a shot list. Every beat carries a frame prompt for z-image and a motion direction for whichever video model you picked.
Features
A brief you can read, images and video that match each other, and a line item for every request that ran.
z-image takes no image input, so a reference is read rather than pasted in. Qwen3-VL writes a precise description of it and folds that into the prompt behind every image in the brief.
Every question the assistant asks comes with the reason it matters, so the next brief you write is better than this one.
How close are we?
Framing decides what the image is about. A macro of one detail and a wide of the whole thing are two different arguments.
A still is a finished deliverable on its own. When you want motion, that exact image becomes the locked first frame. You know what is moving before you pay for it.
Brand name, palette, a default look and a default aspect ratio, saved to your profile once so you are not retyping them into every brief. Small on purpose: fonts, logo files and caption styling are not stored yet.
Queue variants of one brief and 0G Compute routes them upstream in parallel.
0G's broker serves each request over a hardware-attested route. Utsuro writes the model, the work, the cost and that route to 0G Chain as a receipt you can open in the explorer.
| Tx | Model | Work | Cost | Route |
|---|---|---|---|---|
| 0x9f3a…c41b | z-image-turbo | 4 images · 4:5 | 24 cr | |
| 0x1d70…8ae2 | minimax-h3 | 8s motion · 24fps | 240 cr | |
| 0x4c92…07f5 | deepseek-v4-flash | 10k tokens · interview | 2 cr |
The stack
Utsuro does not host GPUs. Every model is reached through 0G Compute on Verified Routing: the route from 0G’s broker to the upstream model is hardware-attested and provable on-chain, and every request settles per call.
Image generation
Photoreal stills at speed. Generate finished images on their own, or use one as the locked first frame of a video so you know exactly what will move.
Video generation
Motion with intent. Takes a prompt, or an image plus a direction, and renders footage with coherent camera work instead of drifting slideshow.
Video generation
The one that arrives with sound. Renders synchronised audio alongside the picture, holds a scene together for longer, and takes a still as its first frame like the other one does.
Creative assistant
The part that makes the other two worth using. It asks what you left out, explains why it matters, and turns a half-formed idea into a prompt worth rendering.
0G Compute
0G Storage
0G Chain
How it works
No format, no keyword soup, no negative prompts. Say what you want and whether you want it still or moving. If you do not say, the assistant asks.
Your idea
A matte ceramic mug for a product page. Four angles.
What you get
Stills and video from the models above. Three of the tiles below are the video files themselves, playing as they came out of MiniMax H3, not frames pulled from them. Each tile carries the model that made it and what it cost in credits.
Pricing
Everyone starts on the same footing: 200 credits, every model, and the whole assistant. Credits are spent per request at the rates below, and the cost is on screen before you press generate.
200
Three published rates and nothing underneath them. No seat fee, no minimum, no per-export charge.
Questions
Still stuck? Write to exorbilabs@gmail.com.
Images and video. Product stills, flat lays, brand portraits, hero frames: z-image renders those on their own, and plenty of people never generate anything else. When you do want motion, two models take a prompt, or a still you already approved, and move it with real camera direction: MiniMax H3, or Seedance 2.5, which renders its own audio. All of it comes out of the same interview.
No, and the gap is large. An image is 6 credits. A second of video is 30, so ten seconds of motion is 300. The assistant session costs a credit or two either way. A four-image product set lands around 26 credits total. A fifteen-second, three-shot video, first frames included, lands near 468.
Not yet. No payment provider is wired up, so nothing in Utsuro can be bought today: no plans, no top-ups, no card field anywhere. Everyone signs up with the same 200 credits and spends them at the published rates. We would rather say that plainly than take pre-orders for a checkout that does not exist.
Not reliably, and we are not going to pretend otherwise. Z-Image-Turbo takes no image input, so a reference cannot condition a still: it is read by a vision model and described into the prompt, which carries subject, palette, framing and mood across a set but not an identity. Saved characters with a locked face are not built. Where a reference does condition the render is video, because both video models accept a real first frame.
You do. Utsuro claims no rights over your briefs, your images, or your finished video. We do not train on your projects, and nothing you make appears anywhere public unless you put it there yourself.
Yes. Commercial use is included, there is no separate licence tier, and there is no watermark on anything you generate. The usual limits still apply: no real person's likeness without their permission, no trademarks you do not own.
Three concrete reasons. 0G Compute reaches every model on Verified Routing: the route from 0G's broker to the upstream is hardware-attested and provable on-chain, so you can check which route served a given request rather than taking a vendor's word for it. What that does not mean: the inference itself runs at the upstream provider, under that provider's policy. The tier attests the path, not the enclave the model ran in, and we are not going to claim otherwise. 0G Storage pins your outputs so they are content-addressed and stay addressable rather than living in a bucket you cannot point at. 0G Chain settles each request at its listed rate, so your bill is a list of receipts instead of a number you have to trust. You never need a wallet.
There are two video models. MiniMax H3 was the first on the 0G Private Computer: text-to-video, or image-to-video from a first frame you have already approved, at 2K, 4 to 15 seconds, billed per generated second. MiniMax's weights went public the same morning it landed, and Artificial Analysis ranked it #2 text-to-video and #3 image-to-video on day one. Seedance 2.5 sits beside it now and renders synchronised audio with the picture, which MiniMax does not, at a slightly higher rate a second. Alongside it, DeepSeek V4 Flash 0731 is now the official release, swapped in under the same deepseek-v4-flash id, with Terminal Bench 2.1 up from 61.8 to 82.7 on a 13B-active MoE. 0G took all three of that morning's launches the same day; the created timestamps are public at router-api.0g.ai/v1/models.
Most of the time you catch it before it costs anything, because you approve the brief first. Rewrite the prompt, change the light, cut a beat, or send the assistant back with a note, and nothing is reserved until you say go. A render already in flight can be cancelled, and a request that fails releases its credits instead of spending them. You only pay for what actually ran.
Start free
One sentence is enough to start. Ninety seconds later you have a brief you would have been proud to write, and images or video to match.