Research-backed calculator guide
AI Outpainting Cost: Budget Aspect-Ratio Extensions Without a Reshoot
Learn how to calculate AI outpainting cost per image across credit, per-run, and per-megapixel billing, then add retries, review, and delivery overhead.
Outpainting adds pixels beyond an image’s original frame. It is useful when one product photo must become a square marketplace image, a wide banner, or a portrait social asset without arranging another shoot. The phrase “AI outpainting cost per image” sounds like it should have one answer, but the bill depends on what a provider counts: a credit, a completed run, a megapixel, a subscription allowance, or a GPU-time estimate.
The practical answer is to calculate two numbers. First, calculate the provider cost per attempt. Then divide the total production spend by the number of extensions that pass review. If a service charges $0.04 per successful outpaint, but 20% of attempts are retried and every accepted result receives two minutes of review, the usable-asset cost is materially higher than $0.04.
Quick answer
For a simple estimate, use:
cost per accepted extension = (attempts × provider cost per attempt + review labor + storage/transfer) ÷ accepted extensions
For a flat-credit API, multiply credits per successful generation by the vendor’s dollar value per credit. Stability AI’s documentation, checked on November 2, 2026, states that outpaint is a flat four credits per successful generation, while its pricing page states that one credit equals $0.01. That makes the documented API charge $0.04 per successful outpaint, before retries, review, storage, taxes, or any application-layer fee. Failed generations are not charged under the outpaint endpoint’s documented rule. [1] [2]
For a hosted model with an approximate run price, keep the estimate separate from a vendor quote. Replicate lists one open model at approximately $0.016 per run, or about 62 runs per dollar, and explicitly says the amount varies with inputs. That is an observed model-page estimate, not a universal outpainting rate or a guarantee for every image size. [3]
The most defensible budget is therefore a range. Use the provider’s published unit for the base case, add your measured retry rate, and report a second number for cost per accepted image after review.
What the provider may charge for
Outpainting is not the same meter everywhere. Before comparing an image extender cost, identify the billing unit and the event that triggers it.
Credits per successful generation. A credit system is easy to model when the provider documents both the credits per operation and the dollar value of a credit. The Stability AI example is four credits for outpaint and $0.01 per credit, giving a $0.04 provider charge per successful generation. The operation requires an input image and at least one non-zero direction such as left, right, up, or down. It also permits a prompt and output settings. [1] This matters because a “single image” may involve several directional changes in your workflow, not just one API call.
Per-run or per-second estimates. A hosted open model may show an approximate run cost. The Replicate model page used for this guide reports about $0.016 per run and says the value changes with inputs. [3] A per-run estimate can be useful for a prototype, but production planning should measure actual billed predictions across the dimensions and masks you intend to send.
Per-megapixel or output-size pricing. Some image APIs price the amount of generated image data rather than the source-file count. In that model, extending a 1,024 × 1,024 image and extending it to 2,048 × 1,024 are not automatically equivalent jobs. Calculate the billed output area using the provider’s stated rounding rule, then multiply by the rate. Do not compare a per-megapixel rate directly with a per-image rate unless the output dimensions are fixed.
Subscription allowance. A web application may bundle generations into a monthly plan. The subscription is not the same as marginal API cost. Divide the portion of the monthly fee allocated to outpainting by the number of usable extensions, and state whether other tools consume the same allowance. If the plan includes unlimited or pooled generations, report an allocated cost rather than calling it a quoted per-image price.
Your own infrastructure. Self-hosting replaces a vendor charge with GPU time, storage, engineering, monitoring, and capacity risk. A low inference number is not automatically a low total cost. Treat hardware, idle capacity, maintenance, and failed jobs as part of the workflow budget.
The official Stability AI announcement describes outpainting as one of a broader set of image generation and editing API services. That positioning is useful for scope: an outpainting estimate should not accidentally include a separate upscale, inpaint, background-removal, or moderation call unless your pipeline actually makes those calls. [4]
A reusable cost-per-image method
Start with the output you want, not the provider’s headline rate. Record the source dimensions, the new target dimensions, the number of sides extended, the model or endpoint, and whether you will generate one candidate or several candidates per input.
Then define an attempt. One attempt can mean one API request, one prediction, or one credit-consuming generation. If an application automatically produces four candidates and a reviewer chooses one, the input has four attempts even when the user sees one final image.
Next, define acceptance. An accepted extension should meet your visible criteria: correct aspect ratio, subject preserved, no obvious seams, no duplicated objects, correct brand details, and an export format that downstream systems accept. A result that renders successfully but is rejected for a warped label is not a usable image.
Use these variables:
u= provider cost per attempt in dollarsn= attempts per source imager= retry or rejection rate, expressed as a decimalh= review minutes per attemptw= fully loaded reviewer cost per hours= storage, transfer, orchestration, and other per-attempt overhead
A first-pass estimate for one accepted image is:
attempts per accepted image = n ÷ (1 - r)
provider cost per accepted image = u × n ÷ (1 - r)
review cost per accepted image = h × w ÷ 60 × n ÷ (1 - r)
total cost per accepted image = (u + s + h × w ÷ 60) × n ÷ (1 - r)
This assumes the rejection rate applies independently to attempts. In practice, a difficult class of images can fail more often than a plain-background product photo, so segment your measurements by image type instead of applying one optimistic average to every SKU.
Worked cost example
Assume a catalog team needs one 1:1 product photo extended to a 4:5 marketplace crop. The team makes one outpaint request at a time, rejects 20% of results after review, spends three minutes reviewing each attempt, and values review time at $30 per hour. Ignore storage and transfer for the base example so the provider and review components stay visible.
For Stability’s documented flat-credit price, u = $0.04. For the Replicate model-page estimate, u = $0.016. The retry factor is 1 ÷ (1 - 0.20) = 1.25, so the workflow expects 1.25 attempts per accepted image.
| Scenario | Input and output assumptions | Provider unit | Attempts per accepted image | Provider cost | Review cost | Total cost per accepted extension |
|---|---|---|---|---|---|---|
| Flat-credit API | One 4:5 extension from one 1:1 source; 20% rejection | 4 credits × $0.01 = $0.04/attempt | 1.25 | $0.050 | 1.25 × 3/60 × $30 = $1.875 | $1.925 |
| Approximate hosted run | Same canvas and review assumptions | Approx. $0.016/run; varies by input | 1.25 | $0.020 | $1.875 | $1.895 |
| Provider-only view | Same as flat-credit case, excluding labor | $0.04/attempt | 1.25 | $0.050 | Not included | $0.050 |
The table shows why a low API price does not necessarily produce a low finished-asset cost. In this scenario, review labor dominates the provider charge. If review is automated and a person checks only exceptions, the labor term may fall. If every source generates four candidates, set n = 4; both provider and review cost rise roughly fourfold before any quality improvement is counted.
The result is a planning estimate, not a quote. The four-credit figure is documented for a successful Stability outpaint call, and the Replicate figure is explicitly approximate. Your own measured acceptance rate should replace the assumed 20% once you have a representative sample.
How canvas size changes the budget
Aspect-ratio extension is a geometry problem. Let the original image be W × H and the target be T_w × T_h. If you preserve the original image without cropping and use outpainting to fill the difference, the added pixel area is approximately:
target area - original area = (T_w × T_h) - (W × H)
That area does not always equal the provider’s billable area. A flat-credit service may charge the same four credits for a small or large permitted extension. A size-sensitive service may round dimensions, enforce maximums, or price the generated area. Your worksheet should therefore carry both physical dimensions and billing dimensions.
For example, extending a 1,000 × 1,000 source to 1,000 × 1,250 adds 250,000 canvas pixels. Extending the same source to 1,600 × 1,000 adds 600,000 pixels. The second job adds 2.4 times as many pixels, even though both can be described casually as “one image.” If the vendor bills per megapixel, that difference belongs in the estimate. If the vendor bills a flat successful generation, record the extra area as a quality, latency, or limit consideration rather than inventing a second fee.
Also account for the number of directions. A single request that expands left and right can be cheaper than two separate requests, but only if the endpoint supports both directions in one call and your workflow accepts the combined result. Stability’s endpoint documents a single request with one or more direction fields, each measured in pixels. [1]
Retries, review, and rejection are part of cost
Outpainting failure is often visual rather than technical. An API can return an image with a warped product logo, an impossible hand, repeated furniture, a visible seam, or a shadow that no longer matches the source. A successful HTTP response is not the same thing as an accepted asset.
Track rejection reasons separately. A prompt adjustment may solve a missing background detail, while a source with a partially visible person may need a different composition or a human decision not to outpaint at all. If you mix these cases together, the average retry rate hides the images that consume the most time.
Use a small review log with the source ID, target ratio, model, attempt number, reviewer decision, failure class, and final accepted attempt. After a pilot, calculate rejected attempts ÷ total attempts and review minutes ÷ accepted images. Those two measures turn a generic image extender cost into a workflow-specific cost.
If a provider does not charge failed generations, that reduces provider waste but does not erase engineering or review time. Stability documents that failed generations are not charged for its outpaint endpoint. [1] You still need to account for request latency, queueing, logging, moderation checks, and any downstream job that runs after a failed response.
Run your scenario
Turn this framework into an editable estimate.
Use dated, source-linked assumptions and see cost per request, successful task and month.
When to use each budgeting approach
Use a provider-only estimate when you are comparing APIs at the earliest prototype stage and have not measured quality. Label it clearly as provider cost per attempt.
Use a cost per accepted extension when an image will be published, sent to a marketplace, or handed to a customer. Include retries and review. This is the better number for a catalog or creative-operations budget.
Use a cost per campaign batch when one source image is adapted to multiple placements. Multiply accepted extensions per source by the number of target formats, then add shared setup and export work. A single square-to-portrait conversion is not the same workload as square, portrait, landscape, banner, and thumbnail variants.
For broader model comparisons, the AI token cost calculator is useful when an orchestration layer adds text prompts, metadata, or captioning. For general API assumptions, compare the LLM cost calculator and the Mistral and Groq API cost calculator, but do not substitute token pricing for an image endpoint’s own meter.
FAQ
Is AI outpainting cheaper than a reshoot?
It can be, but the answer depends on the comparison boundary. Compare the cost of accepted, publishable extensions with the reshoot cost for the same placements. Include creative labor, review, retouching, scheduling, and the risk that an extension fails for a difficult subject. A $0.04 API call is not a complete replacement-cost analysis, and a reshoot may create a better master asset for many future placements.
What is a realistic AI outpainting cost per image?
There is no single universal rate. The documented Stability example is $0.04 per successful outpaint when four credits are valued at $0.01 each. A Replicate model page lists approximately $0.016 per run but says the amount varies with inputs. [2] [3] Your usable cost is higher when you retry, generate multiple candidates, or pay for human review. Report the provider unit and your assumptions beside any number.
Does a larger canvas always cost more?
No. A flat-credit endpoint can charge the same unit for any request within its documented limits, while a size-sensitive endpoint can price output area or round to megapixels. Larger output can still increase latency, rejection risk, storage, and review time even when the provider fee stays flat. Check the endpoint’s billing rule before applying a pixel-area multiplier.
Should I count failed generations?
Count them in the operational workload and acceptance calculation. Whether you count them as provider dollars depends on the vendor’s failure policy. Stability’s outpaint documentation says failed generations are not charged, but the request can still consume time and trigger application work. [1] Keep a separate column for failed attempts, charged attempts, and accepted outputs.
How do I budget a batch of product images?
Start with the number of source images, target placements per source, candidate attempts per placement, rejection rate, review minutes, and reviewer rate. Multiply the provider unit by expected attempts, then add labor and infrastructure overhead. Segment products by complexity; a clean packshot and a lifestyle image with people should not share an untested acceptance assumption.
Is a subscription cheaper than an API?
Only for a defined workload. Allocate the portion of the subscription used by outpainting, divide by accepted outputs, and include any limits on uploads, resolution, exports, or commercial use. Then compare that allocated result with an API estimate using the same number of accepted placements. Do not compare a monthly headline price with an API marginal price without matching the workload.
Assumptions and limitations
This guide uses public documentation and product pages checked on November 2, 2026. Prices, credit policies, model availability, quotas, and plan terms can change. The Stability calculation assumes the documented four-credit successful-generation rate and the pricing page’s $0.01-per-credit conversion. The Replicate comparison is an approximate model-page estimate and varies with inputs. [1] [2] [3]
The worked example assumes one request per candidate, a 20% rejection rate, three minutes of review per attempt, and $30 per reviewer hour. It excludes taxes, storage, CDN transfer, queue infrastructure, prompt authoring, moderation, post-outpaint retouching, and subscription allocation. Replace those assumptions with your own measured values before using the result for a quote or margin decision.
The search benchmark found a mix of tool landing pages, editorial comparisons, community discussion, model catalogs, and API pages. The search interface did not reliably expose sponsored-result labels, so organic classification is provisional. The observed opportunity is a comparable usable-output method, not a claim that any competitor is absent, easy to outrank, or likely to generate traffic.
Primary sources checked
Pricing and product behavior can change. These sources support the time-sensitive claims in this guide; verify current provider terms before committing spend.