GPU & Rendering

1. GPU and Runtime Manager

A runtime is the GPU that renders your scene videos: your own machine's GPU (local) or a remote GPU instance you rent (rented). Rendering a video for the first time on a GPU you manage needs a ready runtime.

Where: Generation assets in the primary sidebar navigation.

A project whose images and audio are already ready can also use subscription-backed Aificient Cloud rendering through Resume Generation, without a ready runtime of your own (see "Aificient Cloud Rendering").

The Generation assets flyout appears beside the sidebar and closes when you select an item, click elsewhere, or press Esc. Inside it:

• The Generation assets row can show compact counts for resources that are actively working and resources that are ready.
• Select the local GPU or a rented instance to open its detail view.
• If cloud jobs exist, the flyout also includes Aificient Cloud.

The manager groups resources by status:

• Running/ready.
• Starting or installing.
• Stopped/inactive.
• Error.

Local Runtime

The local runtime renders videos with your own GPU. It runs the same MiniMax-H3 engine that Aificient Cloud uses, so the picture you get for a given seed is the same; what changes is the time per clip (minutes rather than seconds) and the hardware it asks for.

The model download folder is set in Settings > Model Config.

On a 24 GB card the model does not fit entirely on the GPU, so the engine streams parts of it over PCIe every step. A 5-second clip takes about 4.5 minutes on an RTX 4090 at 720p, and 1080p roughly doubles that (see "The 1080p Warning"). From about 40 GB of VRAM upward the whole model sits on the card.

From the local runtime detail view, you can:

• Install Drivers — install or prepare the runtime.
• Start — start it.
• Stop it.
• Watch setup progress and the live log.
• View logs if something fails (Error — last 12 lines, with Copy logs).

Note: Local and rented renders make clips the same way Aificient Cloud does. Each scene clip is animated from its image with a length derived from the narration (whole seconds between 4 and 15), and the narration is not baked into the clip — it is mixed over it at stitch time.

Messages You May See

Rented GPU Instances

Rented GPU instances are managed from the sidebar and the create-instance modal (see "Renting a GPU"). Only ready instances can be selected for video rendering.

From an instance detail view, depending on state, you may be able to:

• View setup progress.
• View logs in the Instance logs panel.
• Start.
• Stop.
• Reboot.
• Destroy.

The Instance logs panel shows each line's final state (a download progress bar is one line, not every redraw) with colour codes stripped. Copy logs copies that same text and reads Copied once it worked.

When an instance is running, its detail view also shows a render queue for that GPU (see "Render Queue").

Render Queue

Each GPU has its own render queue, and every scene video becomes a job that waits its turn on the GPU you chose for it.

GPUs are not locked to a single project: several projects can line up on the same GPU at once, and a busy GPU is still a valid choice — your videos render after the ones already ahead of them.

The final stitch does not use a GPU: it runs locally in the app once all of a project's scene clips are ready.

Local stitches and caption burns share one local queue that runs one job at a time. If several projects finish rendering around the same time, their stitches run one after another (see "Stitch Node").

The render queue appears in the local runtime and rented instance detail views while the GPU is ready/running. For each GPU it shows:

• Jobs grouped by project, in the order they will run.
• The job currently rendering, with a live progress bar and percentage.
• Queued jobs marked as waiting, with their position in line.

From the render queue you can:

• X on a job's row — remove a single queued video, or cancel the one that is currently rendering.
• Cancel all next to the project name — cancel all of that project's videos across every GPU at once.

Tip: The Generation assets flyout shows a small N queued count next to each local or rented GPU, so you can see at a glance how loaded it is without opening the detail view.

Aificient Cloud Rendering

Eligible projects can render scene videos on Aificient Cloud instead of a GPU you manage. It can appear as a runtime on the initial render-settings screen, and later in the selected project's Resume Generation menu once source images and audio are ready.

Before submission, the Aificient Cloud option shows the project's Lite or Pro tier, resolution, and estimated credit cost. It requires:

• An active subscription and enough credits.
• Each scene's image.
• For scenes whose script has spoken text, the narration generated first, because the narration decides the clip's length.

Aificient Cloud is selected by itself; it cannot be combined with local or rented GPUs in the same submission.

Cost: If your credits do not cover the whole submission, the app sends the scenes one at a time, in order, until the next one is refused for credits. The scenes it could pay for render; the rest are listed in the generation report so you can render them from the scene panel once you have more credits.

How a cloud clip is made:

• The video model animates the scene image and composes the clip's own soundtrack — quiet ambience and sound effects implied by the visible action only, never speech or music.
• Your narration is not baked into the clip; it is mixed over it later, in the preview and in the final stitch.
• Renders made on gift credits carry a small visible Aificient watermark; paid renders do not.

Once submitted, each scene appears as a separate cloud job, and the app refreshes completed clips into the project automatically. If the service answers that some of the scenes are already rendering elsewhere (on your own GPU, or on another device), the submission is retried once without them and nothing is charged for those scenes.

A single queued scene can be taken out of the queue from its video node with Remove from queue, which refunds it, and rendered again later with Resume Generation, even while the rest of the queue is still running.

Note: Poster projects can also render on Aificient Cloud. A poster is a single clip, so the scene-count and scene-length limits above do not apply; it submits one job that carries the project's saved mode, resolution, and duration, and the delivered clip already includes its generated audio.

Cloud Queue and History

When cloud jobs exist, open Generation assets > Aificient Cloud to view the current session's queue and history. The window shows queued, running, completed, failed, and cancelled jobs with progress and status messages.

• Remove an individual queued job from its row.
• Cancel queued — cancels all queued jobs in a project.
• Reserved credits are refunded when a cloud job fails or is cancelled.

While cloud clips are queued or rendering, the center canvas keeps showing the project's overall generation progress, and the bottom of the right scene sidebar shows the Aificient Cloud queue/rendering status. Cancel queued cloud clips from Generation assets > Aificient Cloud; the canvas progress bar does not include a cloud-cancel button.

For account-wide cloud render history, including settled prices, use Settings > Usage > Load latest generation jobs.

2. Renting a GPU

Rent a remote GPU instance when you want to render without using your own card. Every tier renders the same picture; they differ in time and price.

Where: Generation assets, then Create instance (desktop only)

Clicking Create instance opens the Available Instances modal. The web offers Download desktop app instead, with GPU rentals run from the desktop app. If no Vast.ai key is configured, the flyout shows Add Vast.ai key instead.

GPU Tier Cards

The modal header shows the selected tier, its disk, and your price ceiling (for example RTX 5090 · 140 GB disk · up to $3/hr) with a Refresh button. Above the offers table sit the GPU tier cards:

• Four are shown at first — RTX 4090, RTX 5090, RTX PRO 6000, and H100 series. RTX 5090 is the default tier.
• A bare chevron on the rule beneath them (Show 4 more GPUs / Show fewer GPUs) reveals the other four. The list opens expanded when your saved tier is one of the hidden ones.
• Hover a card for a note on what that card does with the MiniMax-H3 model.

Each card shows:

• The tier name and its memory.
• The best live price ($X.XX /hr), or a price hint while offers load.
• Price bar — Higher means lower expected rental cost.
• Speed bar — Higher means faster expected generation throughput.
• Headroom bar — Higher means more of the model stays on the card instead of streaming.

Offers

The search only returns offers that meet all of these:

• At least 23 GB of GPU memory.
• An Ampere or newer card.
• At least 100 GB of disk.
• At least 32 GB of host RAM.
• Your Settings > GPU Config price and reliability limits.

The offers table can include:

• GPU model.
• VRAM.
• Number of GPUs.
• Price.
• Upload and download bandwidth (speed in Mbps and transfer cost per TB).
• Reliability.
• Rent action.

Download Quality Indicator

The Download column shows a small colored status dot next to the speed and cost. It is a quick visual hint for how well an offer can pull large model and asset files, balancing download speed against transfer cost.

Hover the dot to see a short text label. Greener dots generally mean faster instance setup and cheaper data transfer; red dots can mean slower setup or higher costs.

If No Offers Match

The list reads No <tier> offers under $X/hr right now with Try again and Change settings. Change settings opens Settings > GPU Config on top of the modal; the offers refresh when you close Settings.

3. Render Settings

Before starting generation, the app shows render settings. This section describes the History render-settings screen. A poster has its own, shorter version of the same screen — see "Creating a Poster Project" — and a UGC ad has Video settings — see "Video Settings for a UGC Ad".

Format

• Aspect ratio — Portrait (9:16) (default), "TikTok, Reels and Shorts", or Landscape (16:9), "YouTube and desktop". Use 9:16 for vertical/social video and 16:9 for widescreen video.
• Resolution — 720p (default) or 1080p.

Cost: 720p is much cheaper to render with very similar quality; 1080p costs more per clip.

Runtime

For the first generation, choose one or more ready GPUs, or select Aificient Cloud if it is available for your account.

• Each ready local or rented runtime is listed with a checkbox; tick every GPU you want this project to use. Each option shows how many jobs are already in queue on it.
• By default the app pre-selects the lightest-loaded GPU.
• Aificient Cloud is a separate, mutually exclusive target: selecting it clears the local/rented GPU selection. Its row shows the selected Lite or Pro tier, resolution, and estimated credit cost.

What each choice does:

• One GPU selected: all scene videos render on that GPU, in queue order.
• Several GPUs selected: the scene videos are split across the chosen GPUs so the project finishes faster.
• Aificient Cloud selected: scene videos are submitted to the cloud after their source images and audio are prepared; the app continues through captions and final stitching as the clips arrive.

The Action Button

The button at the bottom of the screen reads:

• Start generation — the default.
• Start images and audio — when Images and audio only is on.
• Queue on N GPUs — when several GPUs are selected.

It comes with a summary line of the chosen aspect, resolution, tier, and target.

Messages You May See

Quality

Two tiers are shown as labelled buttons, each with an icon and a short blurb:

• Lite — "Faster, lower VRAM": renders faster and runs on smaller GPUs.
• Pro — "Best quality": takes longer but gives the best quality.

The default depends on your account: trial accounts (welcome credits, no plan) start on Lite, and accounts on a paid plan start on Pro. You can switch either way before starting the generation.

Suggested use:

• Lite for previews and quick feedback, and on a 24 GB card when time matters.
• Pro for finished videos.

On every runtime the tier is a step-count knob, not a precision setting: Lite renders each clip with a few fast steps, Pro with many more. Any tier can run Lite, including an A100.

The engine decides how much of the model to keep on the card from its VRAM, so the picture for a given seed is the same on every GPU — only the time changes.

Cost: On Aificient Cloud, Pro uses about a third more compute than Lite and is priced accordingly.

Messages You May See

Images and Audio Only

Enable this option when you want to generate only:

• Character references.
• Scene images.
• Narration.
• Sound effects.

Video rendering and final stitching are skipped.

Tip: Use it to review assets before spending GPU time.

4. Resume Generation

Resume Generation is the button at the bottom of the right sidebar. It lights up amber when a project has missing or invalidated assets, and is disabled (Project generation is complete) when there is nothing left to do. The app reuses existing completed assets where possible and only regenerates missing or affected parts.

When to Use It

Use Resume Generation after:

• Stopping generation.
• Deleting an image, audio, or video.
• Rewriting scene text.
• Editing character visuals, changing a scene's cast, or changing an outfit or a scene's assets.
• Adding a character or an asset.
• A generation failure.
• Running images/audio first and rendering video later.
• Removing a queued clip from a GPU or cloud queue, or cancelling one cloud scene while the others render.

Note: A new character that is not in any scene yet also counts: its reference image is missing, so the button stays active even on a project whose final video is already finished, and Images + Audio only renders it for preview without touching the video.

The Generate video on Menu

Clicking the button opens the Generate video on menu, which contains:

• Generation changes pending to be saved! — a warning shown when the Settings tab has unapplied changes. The render action stays blocked until you apply or discard them.
• The runtime list — Aificient Cloud (with tier, resolution, and estimated credits) and each ready local or rented GPU (with its queue depth). Tick one or more GPUs to split the missing videos across them, or Aificient Cloud by itself. When nothing is available the list reads No GPU available — start a runtime to render video.
• Add to render queue (or Queue on N GPUs) — renders the missing scene videos on the selected runtime(s). Missing images and audio are generated first automatically.
• Images + Audio only / Skip video rendering — generates missing images, character references, project assets, and audio without rendering video. It stays available while the cloud queue is busy, so a failed narration can be regenerated while clips render.

While images or narration are still missing, the Aificient Cloud row reads Lite · 720p · generates the missing images/audio first (or Pro · …) instead of a price. Picking it generates those assets first and queues the clips only if every one of them succeeds.

Note: Like a GPU run, a failed asset stops before any clip is paid for, and the failure is in the generation report.

When every raw clip is ready and only caption burns or the final stitch remain, the button acts directly instead of opening the menu: it runs the remaining caption burns and final stitch locally, and no GPU is needed for that work. If only the final stitch is left, it is labelled Stitch Final Video.

On a phone the button is a floating Resume / Stitch pill beside the Details pill, and the menu opens as a bottom sheet. While clips render on Aificient Cloud, the floating Cloud pill (it opens the render queue) shrinks to its icon and count and moves left so the Resume pill fits beside it.

On a poster project the same menu offers a single runtime choice plus Images only — see "Rendering and Resuming a Poster".

Messages You May See

Resume Beside a Live Cloud Queue

A busy Aificient Cloud queue does not hide Resume Generation. When at least one scene has no clip and is not in the queue — because you removed it, cancelled it, or deleted its clip — the button stays available. Its tooltip reads N scene videos are missing and not in the cloud queue — render them while the queue runs.

The opened menu notes Renders only the N missing scene videos. The scenes already in the cloud queue keep rendering there; the final video is stitched once every clip is in.

The resumed run renders only those scenes, on the cloud or the GPUs you tick. It never pays twice for a scene already queued, waits for the cloud clips, and then runs the captions and the stitch itself.

Note: If a cloud job you were waiting on ends without a clip, that scene's video phase reports The Aificient Cloud render ended without a clip.

Generating on Another Device

Every client — this app, the web, the iPhone app — registers the generation runs it starts, so the others can show them. Each reports its progress within each step every few seconds, so the percentage moves smoothly; the iPhone app also follows the run scene by scene on the lock screen.

If a device crashes mid-run, the scenes it held are freed within about two minutes.

While another device is generating the open project, the sidebar footer shows a sky-blue chip instead of the generate controls: Generating on <place> N%. The place is <GPU name> · <device>, a rented <GPU> · <device>, or Aificient Cloud · <device>, with a phone, cloud, or chip icon.

The project row in the left sidebar blinks like any other running project.

The chip's tooltip adds the current stage and a new generation of this project waits until it is done. The stages are:

• Drawing the images
• Recording the voices
• Rendering your clips
• Putting your video together

Note: There is nothing to stop here: the device running it is the one that stops it. Resume Generation is hidden meanwhile, and starting a run anyway is refused with This project is being generated on <place>.

Nothing about other devices' runs is stored on this one; the list is rebuilt from the server every time the app starts.

The 1080p Warning

When you queue scene videos at 1080p onto a GPU that has to stream the model — any card under 40 GB of VRAM, which covers every 24–32 GB card, local or rented — the app asks first. The dialog is titled 1080p will be slow on <card> (several cards are joined as A, B and C).

It explains: It cannot hold the whole model, so every step pulls it back over PCIe — and 1080p renders twice the pixels. Expect clips to take up to twice as long, and anything past ~6 seconds not to fit at all.

Your choices:

• Switch to 720p (Switching…) — saves the project's resolution as 720p and then queues the same GPU selection. If the save fails nothing is queued.
• Keep 1080p — queues the render as selected.
• Clicking outside the dialog — cancels and queues nothing.
• Don't ask again — stops the dialog for good on this computer.

Warning: Don't ask again applies to every account on this computer, and there is no setting to turn the dialog back on.

Note: The dialog never appears for Aificient Cloud, for Images + Audio only, or for cards from 40 GB upward (L40S, A100, RTX PRO 6000, H100, H200).

What Resume Can Run

If the project already has images and audio, Resume moves directly toward video rendering once you pick Aificient Cloud or one or more of your own GPUs; if only video is missing, the app does not need to recreate the whole concept. Work sent to your GPUs is added to their render queues, so it starts even if a GPU is currently busy.

5. Stop Generation

When generation is active, the sidebar shows a stop/cancel action. You can usually continue later with Resume Generation.

Stopping generation:

• Stops the current app-side generation flow.
• Attempts to cancel active video work.
• Marks unfinished phases as cancelled.

Note: This stop action is for video and asset generation. To cancel a brainstorm, story, or character request on the creation home screen, use the stop button that replaces the chat's send button (see "Stopping an In-Progress Chat Request").

6. Generation Error Summary

When a generation finishes and one or more parts of it failed, the canvas shows an error summary box in the top-right corner. If everything succeeds, no box is shown.

• The box starts collapsed, showing only a header such as Generation finished with 3 errors.
• Click the header (or its chevron) to expand the full list, and click it again to collapse it.
• Click the trash/Clear button next to the header to dismiss the summary for the current project.
• The box stays on screen until you clear it, or until you start a new generation or resume the project.
• The box reports results for the currently selected project. Switching projects hides it; it reappears for the project it belongs to until you dismiss it or start a new run.

What It Shows

When expanded, the box lists every failure from the most recent finished generation, grouped into two categories:

• Runtime tasks — work that runs on the GPU runtime (video generation and audio alignment) plus the final stitch, which runs locally in the app.
• API tasks — work that runs on the cloud service, such as character images, scene images, narration, and sound effects.

Each entry shows:

• The type of generation that failed (for example, Video generation or Audio).
• The scene it belongs to, when the failure is scene-specific.
• The reason reported for the failure.

If a reason is long, it is shortened with a Show more link. Click Show more to read the full text and Show less to collapse it again.

Note: Cancelled steps are not listed as errors. If you stop a generation, the box still reports any real failures that happened before you stopped, but not the steps that were simply cancelled.

What To Do With It

Use the summary to decide what to fix or rerun:

• API task failures (images, audio, characters) — review or rewrite the affected scene or character, then use Resume Generation.
• Runtime task failures (video, alignment) — check that a GPU runtime is ready and review its logs, then use Resume Generation.
• Stitch failures — these happen locally. Check your internet connection and disk space, then use Resume Generation to retry the stitch.
• A 409 error — the runtime is busy with a task, possibly due to a state mismatch between the runtime and the UI. Wait a few minutes and try again; if the issue persists, restart the runtime.