Runware
AIGenerate images and video, upscale, and remove backgrounds across hundreds of models through Runware's fast, low-cost inference. Agents produce visual assets at volume inside the workflow that needs them.
What This Integration Enables
Runware is the production line in this category, and the pages that surround it are the studio. Its argument is throughput and unit cost across a catalogue of hundreds of models through one endpoint, which makes it the right choice when the question is not whether a single image is exquisite but whether 300 of them can exist by Thursday without a budget conversation.
Two details make it usable at that scale rather than merely cheap. First, generation exposes real control: dimensions, sampling steps, CFG scale, scheduler, and seed, plus image-to-image with a strength value. Seed control is what turns a good result into a repeatable one, so a flow can record the seed alongside the asset and reproduce that render later instead of rolling the dice again. Second, delivery is a choice per request. Fast work returns inline; slow work, video especially, sets Delivery Method to Asynchronous and is collected with Get Task Result polled on exponential backoff rather than by holding a connection open for minutes.
The part worth planning for before you build is retention. Results come back as hosted URLs kept for seven days by default, and Time To Live changes that. Seven days is enough for a review cycle and not enough for an archive, which means the flow needs an explicit decision about what gets copied into durable storage such as S3 and what is allowed to expire. Get Account Details is the other half of running this responsibly: it returns the credit balance, the team roster with roles, every API key with its lifetime request count, and rolling usage for today, the last 7 days, the last 30 days, and all time. That is enough to attribute spend to a project rather than to a mystery.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Catalog imagery at batch scale
A seasonal refresh needs lifestyle imagery for hundreds of items. The agent builds a prompt per item from the product attributes, submits the batch through Generate Image with a fixed seed per item so the result is reproducible, screens the outputs on dimensions and completeness, and posts a contact sheet for review. Approved images are copied into durable storage and written to the product records in Shopify. The unapproved ones expire on their own.
Product cut-outs with clean edges
Supplier photography arrives on inconsistent backgrounds. The agent runs Remove Image Background with alpha matting enabled, which is the setting that decides whether hair and fine detail survive the cut, sets the output background as an RGBA value to match the storefront, and runs Upscale Image on anything below the listing resolution using detail and realism enhancement. The cleaned set is filed against the listing.
Short product video without a held-open connection
A launch needs a few seconds of motion per hero item. The agent submits Generate Video with Delivery Method set to Asynchronous, receives a task reference, and polls Get Task Result with exponential backoff while the rest of the flow continues doing useful work. When the status reaches success, the hosted MP4 URL is collected, copied into durable storage, and attached to the campaign brief in Notion.
Human-in-Loop Highlight
The seven-day clock on hosted results is what forces a person into this flow, and it is the right thing to force. A batch run produces far more images than anyone will publish, every one of them sitting on a URL that stops working in a week, and copying all of them into permanent storage is just moving the waste. So the agent narrows and asks. It screens out the technical failures, groups the survivors by product, posts a contact sheet to the merchandiser in Slack, and says: "284 images passed screening across 300 items. The hosted URLs expire in six days. Select the ones that go live and I will copy those into permanent storage with their seeds recorded. Anything not selected expires, and regenerating it later means paying for the run again." The person picks. The agent persists exactly that set, keeps the seed with each one so the render can be repeated, writes them to the listings, and lets the rest go. Taste is not something to automate, and neither is deciding what a brand keeps. The agent's job is to make the choice small enough that a person can actually make it. That is what Orchestration as a Service looks like at the level of a single afternoon's work.
Agent Capabilities
7 actionsGeneration
2- Generate Image Generates one or more images from a text prompt using any image model in the catalog, with image-to-image supported by supplying a seed image and a strength value. Exposes dimensions, sampling steps, CFG scale, scheduler, and seed, which is what makes a result reproducible. Returns hosted URLs retained for seven days by default, adjustable with Time To Live. For slow models, set Delivery Method to Asynchronous and collect with Get Task Result.
- Generate Video Generates a video from a text prompt, and optionally from frame images, using any video model in the catalog. Video generation is slow, so set Delivery Method to Asynchronous and poll rather than holding the connection open. Returns a hosted video URL retained for seven days by default.
Image Editing
2- Upscale Image Increases the resolution of an existing image using an upscale model. Set either an upscale factor or a target megapixel count, never both, since the two are mutually exclusive. Detail and realism enhancement improve fine textures and are particularly effective on generated images.
- Remove Image Background Removes the background and returns the cut-out subject. Alpha matting refines edges around hair and fine detail, the output background can be set as an RGBA value, and the mask alone can be returned instead of the composited image when a later step needs to do its own compositing.
Tasks
1- Get Task Result Retrieves the current status and any available results for a task submitted with Delivery Method set to Asynchronous, returning processing, success, or error along with a progress percentage where the model reports one. Poll with exponential backoff starting at one to two seconds rather than in a tight loop.
Models
1- Search Models Searches the catalogue across featured, community, and private models, matching on names, versions, and tags, and narrowing by category, architecture, and visibility. Returns paginated results whose AIR identifiers pass straight into the generation operations.
Account
1- Get Account Details Returns a point-in-time picture of the organisation behind the configured key: the current credit balance, the team roster with roles, every API key with its lifetime request count, and rolling usage totals for today, the last 7 days, the last 30 days, and all time. Used for billing automation, project-level spend attribution, and low-balance alerting.
Frequently Asked Questions
What can FlowRunner do with Runware?
FlowRunner agents can run Generate Image, Generate Video, and Upscale Image in Runware, plus 4 more actions.
Does connecting Runware to FlowRunner require OAuth?
No. Runware connects to FlowRunner with an API key, no OAuth flow required.
Can Runware trigger a FlowRunner workflow automatically?
Runware doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with Runware
$100 in credits. No card required. Connect in minutes.