fal vs Replicate: When is a batch of generated product assets really delivered?

For teams sending generated images and videos to their own asset library: fal’s queue and media-retention controls suit explicit handoff windows, while Replicate’s event filters suit existing prediction pipelines. Both require independent file storage and a distinction between generation success and business acceptance.

Source: fal.ai
November 8, 2026
5 min

When a marketing-product team generates product images and short videos in batches, it needs assets that people can find, review, and publish. Both fal and Replicate provide API access to hosted models. First establish whether the target model can represent the products adequately, then compare how the resulting files reach your own system.

If you need per-request media-retention periods, evaluate fal first. If you already have a system organized around predictions and want notifications for starts, outputs, logs, or completion, Replicate fits that structure more readily. Neither makes a successful callback a substitute for asset ingestion and business approval. These judgments draw on official documentation, not account testing or a ranking of generation quality.

Separate “finished” from “usable”

fal’s queue submission endpoint returns a request_id plus status, result, and cancellation URLs. After the status reaches COMPLETED, you still need to retrieve the result. The result endpoint passes through model errors, so queue completion does not mean successful generation. Replicate calls each invocation a prediction and distinguishes starting, processing, succeeded, failed, canceled, and aborted. An aborted prediction exceeded its deadline before it began running.

For an asset library, both state models ultimately need mapping to your own records: submitted, generation succeeded, file saved, awaiting review, and approved. Keep the business batch ID, product identifier, platform job ID, model identifier, and generation parameters so retries do not mix up versions. This is a delivery-process design, not an acceptance feature supplied automatically by either platform.

Record the outcome of cancellation too. fal removes queued requests; for running requests, it signals cancellation and leaves the model app to decide how to stop. CANCELLATION_REQUESTED does not mean execution has stopped. Replicate provides an endpoint to cancel a running prediction. With either integration, continue checking the final state and prevent late results from a withdrawn batch from automatically entering publication.

Callbacks reduce waiting, but do not remove reconciliation

fal sends a webhook when an asynchronous job ends. Its documentation requires a 2xx acknowledgment. Network errors, timeouts, and 4xx or 5xx responses trigger backoff retries, up to 31 times, bounded by result retention: usually about 1 hour, or about 6 minutes for results of 10 KB or more. A 3xx redirect is not retried. That size refers to the API result, not the generated video file.

Replicate lets you select notification types with webhook_events_filter. A completed event can also mean failure or cancellation; do not ingest merely because you see completed. Its receiving guide says failed terminal callbacks are retried with backoff, with the final attempt about 1 minute after completion. Intermediate states are not retried, and notifications can be duplicated or arrive out of order. This offers more flexibility for progress notifications, but adds state-merging logic.

For both, verify the signature, persist the event, and acknowledge promptly, leaving file downloads to background processing. Add reconciliation jobs to recover results whose notifications were missed. fal verifies ED25519 signatures using JWKS public keys; Replicate uses a user- or organization-specific signing key and HMAC-SHA256. Both require verification against the raw request body and timestamp checks; reserialized JSON cannot replace the original body. You still need idempotency keyed by the platform job ID to prevent duplicate ingestion.

Save files to your own storage first

fal’s retention documentation separates CDN files from request records. Generated files are returned as URLs. X-Fal-Object-Lifecycle-Preference can set expiration_duration_seconds; null means no expiration, and a per-request setting overrides the account default. Request inputs and outputs are retained for 90 days by default, which does not mean media files must remain downloadable for 90 days. Expired files are permanently deleted, so define your download window.

Replicate API predictions have their input parameters, output values, output files, and logs removed after 1 hour by default. Web-interface runs follow different retention rules; website history does not establish API file lifetime. The client’s replicate.run() can return FileOutput objects, while the HTTP API represents output files as HTTPS URLs. Whether you receive an object or a link, continued use requires saving a copy.

Suppose a team generates short videos overnight and reviews them the following morning. With either service, automatically download the files to owned storage after generation, then give the review page persistent addresses. Replicate’s default 1-hour cleanup means this cannot wait for staff to start work. fal’s configurable retention gives more control, but still requires explicit settings and verified downloads rather than treating vendor links as permanent library addresses.

Choose using accepted assets after ingestion

Use a small, representative set of product briefs to validate the integration: normal completion, generation failure, cancellation during processing, duplicate callbacks, a temporarily unreachable callback endpoint, and failed file downloads should all produce traceable outcomes. Check product appearance, text, aspect ratio, and video duration before counting approved files as delivered. This checklist is a proposed trial method, not a report of tests already performed.

If both target models meet your needs, teams prioritizing per-request media-retention policies can trial fal first; teams prioritizing multiple progress events and an existing prediction data structure can trial Replicate first. Without automated downloads, reconciliation, and review records, neither is suitable for directly handling unattended batch delivery. Assess budget as total generation and retry costs plus storage and processing costs, divided by the number of approved assets. Identical model names do not establish equivalent execution settings, quality, or cost per usable asset.

Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.

Original source:fal.ai