Learn the craft,
not the interface.
Walkthroughs written against what the composer actually accepts today — real limits, real rates, every step printed on the card.
Three steps to your first clip
One prompt, one settings panel, one wait. All of it works anonymously — no account and no card. 200 credits is the allowance we quote, not a balance a server keeps for you.
Write the prompt
Say what the subject is, what it does, where it stands and how it is lit. “Try a prompt” loads a working example to edit instead of an empty field, and “Add cinematic direction” appends lighting, depth of field and camera language for you.
Pick the model and settings
The panel only offers what the chosen model really supports — resolutions, aspect ratios, lengths, audio. The cost line moves as you change them, so the reserved credits are visible before you commit to anything.
Generate, then continue
The dialog reports elapsed time and an ETA and keeps running if you send it to the background. When the clip lands, download it before the link expires in about 24 hours — or chain straight into the next shot from its last frame.
Walkthroughs, in full
Every step is on the card. Nothing hides behind a detail page, and there is no video to sit through.
Write a prompt the model can follow
Four clauses in a fixed order beat a paragraph. Prompts are capped at 2000 characters, and the panel warns past 500.
- Open Create in the workbench and, if a blank field is intimidating, load an example with “Try a prompt” and edit it rather than starting from nothing.
- Write one clause each for subject, action, setting and light, in that order — “neon koi swimming through a flooded Tokyo alley, rain-lit, handheld”.
- Put the look last: film stock, lens, time of day. “Add cinematic direction” appends a tested lighting-and-camera tail if you would rather not write one.
- Stay under 500 characters on Vgent 1.0; longer prompts still submit but get harder for the model to hold, and anything past 2000 is rejected outright.
- Change one clause between runs. Move two and you learn nothing about which one helped.
Reference images, start frames and end frames
Image mode drives from a start frame; reference mode carries a subject between shots. Uploads are not connected yet, so plan around the frames you already own.
- Prepare images inside the shared envelope: JPG, PNG or WebP, 300–6000 px per side, width-to-height between 0.4 and 2.5, under 30 MB — 10 MB for Veo 3.1, Kling 3.0 and Wan 3.0.
- Drop a file on the composer: it is probed and checked in the browser and names the exact limit it misses. Object storage is not wired up yet, so the file cannot be sent — the check is there so your assets are ready when it is.
- Know what each mode wants. Vgent 1.0 takes a start frame with an optional end frame; Wan 3.0 image mode requires both; Veo 3.1 and Kling 3.0 accept one or two.
- Use reference mode to hold a character, product or location steady across shots — up to 9 images on Vgent 1.0, 3 on Veo 3.1.
- The one image path that works end to end today: finish a Vgent 1.0 or Seedance 2.5 clip, then press “Continue from last frame” to reuse that frame as the next start frame.
Camera motion that survives the render
No model exposes a camera parameter — every move is written in the prompt. One move per shot, and say where it ends.
- Name the shot size first — wide, medium, close — so the model has a frame to move away from.
- Add exactly one move: slow push in, orbit left, crane up, handheld follow. Two moves inside five seconds usually delivers neither.
- Say where the move stops — “push in until the face fills the frame”. An open-ended move drifts and the last second wanders.
- Match the move to the length: 5 seconds holds a single beat, ten carries a move plus a reveal, and Seedance 2.5 or Wan 3.0 will run to 30.
- Iterate at 480P while the motion is still wrong, then re-run only the winning prompt at 1080P.
Audio: what each model gives you
Native sound is per-model — some always include it, one bills for it, one has none. Lip sync and voice cloning are not connected yet.
- Vgent 1.0, Seedance 2.5 and Wan 3.0 generate audio alongside the picture. The toggle is on by default and does not change the price.
- Kling 3.0 bills sound separately: 720P goes from 28 to 40 provider credits a second, about ⚡45 more on a 5-second clip.
- Veo 3.1 always returns audio, music included. There is no toggle, and it is priced per clip rather than per second.
- MiniMax H3 Max has no audio at all — the control disappears from the panel instead of failing at submit time.
- Direct the sound in the prompt: “rain on a metal roof, distant traffic”. Lip sync, dubbing and voice cloning live in Virtual Human and Translator, which are not connected yet.
Choose resolution and length for the credits you have
Most models bill per output second, so length and resolution are the two dials that matter most. 200 credits — about one finished 720P clip — is a quoted allowance, not a balance a server holds.
- Draft at 480P. On Vgent 1.0 that is 18 provider credits a second against 38 at 720P and 95 at 1080P — the same prompt, a quarter of the cost.
- Read the cost line before submitting. It shows what will be reserved, with Vgent credits calculated as ceil(provider credits × 0.75).
- A 5-second 720P clip on Vgent 1.0 costs ⚡143; the identical clip at 1080P costs ⚡357. Resolution, not length, is usually what breaks a budget.
- Remember that reference video is billed too on Vgent 1.0, Seedance 2.5 and Wan 3.0: input seconds are added to output seconds before the per-second rate applies.
- What is actually enforced: a cap on each clip’s estimate — at most 2 images and no video or audio references — plus 5 submissions per 10 minutes per network address and one daily budget shared by every anonymous visitor. If Generate refuses, the panel names the limit you crossed.
Continue a shot and extend a clip
Two different chains: last-frame continuation on Vgent 1.0 and Seedance 2.5, clip extension on Veo 3.1. Both are short-lived.
- When a Vgent 1.0 or Seedance 2.5 clip finishes, press “Continue from last frame”. The composer switches to image mode with that frame already loaded as the start frame.
- Write the follow-up as a continuation of the action, not a new scene — the frame already fixes subject, framing and light, so only the movement needs describing.
- The chain is signed on the server and only valid for the model that produced it. Switch models and the panel drops it and tells you why.
- Veo 3.1 offers “Extend this clip” instead, and only for 720P sources. Pick fast, quality or lite — each option shows its price on the button before you commit.
- Save the frame you care about: sources and output links both expire roughly 24 hours after the task completes, and the chain goes with them.
Work the library instead of your downloads folder
The library is this browser’s record of the last 24 hours: filters, sorting and density, with no account behind it.
- Open Library in the workbench sidebar. Every task from this browser is listed with its model, settings and the credits that were reserved.
- Narrow the list by kind, by model and by status — in progress, completed, failed — then sort newest or oldest first.
- Switch grid density when you are scanning many results rather than studying one.
- Download anything worth keeping the same day. Provider links expire in about 24 hours and the file is not mirrored anywhere.
- There is no account yet, so clearing site data or moving to another browser loses the list. Treat it as a working surface, not an archive.
Read a failure before you retry it
The dialog separates a rejected request, a busy provider and an uncertain submission. Each one wants a different move, and only one of them is “try again”.
- Content filter: reword the prompt or swap the reference. Celebrities and public figures in a reference image are refused by the provider’s moderation, and the task fails before it renders.
- Model offline or provider busy: switch models, or wait out the countdown shown on the retry button instead of resubmitting into the same queue.
- Uncertain submission — a timeout or a dropped connection — keeps waiting on the task that already exists. Do not create a replacement; you would pay for both.
- Failed tasks show a refund line. Where the provider still reports a charge, it is flagged for reconciliation rather than promising credits back that nobody has confirmed.
- Anonymous limits are counted per network address, not per browser: 5 video requests per 10 minutes, on top of a daily budget shared by everyone. On a shared or office network, the people around you spend the same allowance. Both limits say how long to wait.
Where to go from here
Reference material, the endpoints behind the panel, and the people using them.
Documentation
Model-by-model limits, parameter ranges and what each control in the panel maps to in the request.
API reference
Request shape, task lifecycle and error codes. The API serves this site today; public keys are not open yet.
Community
Where prompts, settings and finished clips will be shared once the channels open.
Every rate, format and cap above is read from the same descriptors the composer validates against, so a guide and the panel cannot disagree. Providers move fast, and the panel moves with them.