AI image generation costs depend on more than the advertised price of a single image. Resolution, quality mode, retries, batch size, storage, moderation, subscriptions, and local compute can all change the final budget. This guide provides a repeatable way to estimate the cost of generating AI images for a campaign, compare hosted APIs with subscriptions and local workflows, and identify which assumptions need to be checked before production.
Overview
The simplest cost question is: how much does one generated image cost? The more useful question is: how much will it cost to produce, select, prepare, store, and deliver the usable images a project needs?
An image model may charge by image, resolution, quality setting, compute time, credits, or another usage unit. Some tools bundle generation into a monthly subscription, while an API usually makes usage easier to measure directly. A local workflow may avoid per-image API charges but still has hardware, electricity, setup, maintenance, and operator costs.
For a reliable estimate, separate these categories:
- Generation: the model calls or credits used to create candidates.
- Iteration: retries, variations, upscaling, inpainting, and outpainting.
- Production: background removal, editing, formatting, and quality review.
- Delivery: storage, bandwidth, moderation, and application infrastructure.
- Fixed access: subscriptions, seats, minimum commitments, or local equipment.
Do not compare providers using only the lowest listed generation price. A cheaper first draft can become more expensive if it requires more retries, has limited resolution options, or does not fit your application workflow. For broader capability comparisons, see the guide to text-to-image models, prompt control, and output quality.
How to estimate AI image generation cost
Use this basic formula for a campaign:
Total estimated cost = generation cost + iteration cost + processing cost + storage and delivery cost + fixed access cost.
Start with the number of final images, then work backward from your expected selection rate. If a team needs 100 approved images and typically keeps one image from every five candidates, the working estimate is 500 generations before adding a contingency. The selection rate should come from a small test batch whenever possible rather than an optimistic assumption.
A practical worksheet can use these fields:
- Final approved images required.
- Average candidates generated per approved image.
- Number of campaigns, formats, or aspect ratios.
- Percentage of images requiring a second pass.
- Resolution and quality tier for drafts and final outputs.
- Unit price or credit consumption for each operation.
- Storage, delivery, and other per-use charges.
- Contingency for failed calls, prompt changes, and review revisions.
Then calculate separately for drafts and finals. Low-cost draft generation may be suitable for composition checks, while final assets may require a higher quality setting or an additional upscale. Treating every call as a final-quality call can overstate the budget; treating every call as a draft can understate it.
Compare pricing structures, not just headline rates
| Pricing structure | Useful for | Cost variable to record | Potential blind spot |
|---|---|---|---|
| Pay per API generation | Products, automation, and predictable batch jobs | Model, resolution, quality, and number of calls | Retries, storage, moderation, and failed requests |
| Credits or subscription | Individual creators and recurring manual work | Monthly fee, included credits, and overage terms | Unused allowance, queue limits, or restrictions on commercial workflows |
| Local inference | Teams needing control over data or repeated usage | Hardware, electricity, setup, and maintenance | Operator time, model setup, and slower throughput |
| Hybrid workflow | Drafting locally and finalizing through a hosted service | Costs for each stage and transfer between systems | Extra implementation and asset-management complexity |
Before selecting a provider, record the current price, billing unit, supported resolutions, quality modes, rate limits, and subscription terms in your worksheet. Prices and model lineups change, so the worksheet should include a date and a link to the provider's official pricing page.
Inputs and assumptions
Resolution is one of the most important inputs. A square thumbnail, a landscape editorial image, and a large poster may use different tiers or operations. List each required output format instead of using one average resolution for the entire campaign.
Quality settings also affect the estimate. Use a draft setting for layout exploration only when the output can be discarded or replaced. If every candidate must be delivered, draft savings may not apply. Likewise, count image editing operations separately when an API charges for edits, variations, masks, or upscaling.
Prompt quality changes cost indirectly. Clear subject, composition, style, lighting, and aspect-ratio instructions can reduce retries, but no prompt guarantees a successful first result. A consistent prompt template and a defined review checklist make the selection rate easier to measure. The prompt template guide can help teams standardize inputs before calculating production volume.
Include non-generation costs when they are part of the workflow:
- API request handling and queue infrastructure.
- Object storage and image delivery.
- Content moderation or safety review.
- Human review and correction time.
- Image conversion, metadata handling, and accessibility preparation.
For local models, estimate the portion of equipment cost assigned to the project rather than pretending the hardware is free. A simple allocation can divide the equipment's planned useful hours by the hours used for this campaign. For implementation details, see how to run text-to-image models locally.
Worked examples
Example 1: a fixed campaign of approved images
Suppose a publisher needs 40 approved illustrations. Its test batch suggests that one approved image requires four candidates, and 20% of approved images need a second generation pass. The estimated generation volume is 40 × 4 = 160 initial candidates, plus 8 additional passes, for 168 image operations. Add the applicable unit cost for the selected model and quality tier, then add editing, storage, and review time.
This approach is more useful than multiplying 40 by a single advertised image price because it reflects the actual selection process. If the campaign uses two aspect ratios, calculate the split separately; some prompts may need to be regenerated rather than cropped.
Example 2: subscription versus API
A creator compares a monthly subscription with API billing. The subscription is easier for manual exploration, but its value depends on how much of the included allowance is actually used and whether the terms fit the intended workflow. The API may be preferable for an application because each request can be logged and assigned to a user, project, or content item.
To compare them, calculate the monthly cost at three volumes: low, expected, and high. Include overage or additional access requirements where applicable. The break-even point is not universal; it depends on the provider's current terms, the team's usage pattern, and the amount of automation required.
Example 3: a recurring content pipeline
For weekly thumbnails or social assets, estimate a month rather than a single deliverable. Multiply the number of publishing cycles by the number of assets per cycle, then apply the observed candidate-to-approval ratio. Track actual calls and approved outputs for the first few cycles. Replace the initial assumption with the measured ratio when you renew the forecast.
A repeatable pipeline is easier to budget when prompts, model settings, dimensions, and review rules are versioned. The repeatable AI image workflow guide covers the operational side of this process.
When to recalculate
Recalculate your AI image generation cost whenever a provider changes model pricing, resolution tiers, quality modes, credit rules, rate limits, or subscription terms. Also revisit the estimate when a model is replaced, a new image format is introduced, or your workflow moves from manual creation to API automation.
Operational changes matter too. Recalculate after a prompt template revision, a change in approval standards, a new aspect ratio, or a shift from draft-first generation to final-quality generation. If the team begins using image editing, upscaling, moderation, or automated storage, add those operations to the worksheet rather than hiding them inside a general contingency.
At the end of each campaign, compare three numbers: planned generations, actual generations, and approved images. Record why the numbers differed. Common causes include prompt failures, unsuitable compositions, format changes, and review revisions. Use those observations to update the candidate-to-approval ratio and the contingency.
For a practical next step, create a dated cost sheet with one row per model and output tier. Run a small test batch, measure the number of candidates needed for an approved image, and forecast low, expected, and high usage. Recheck provider pricing before committing to production, and link the worksheet to your API implementation or content workflow. This turns AI art generation costs from a vague per-image question into a decision that can be reviewed and recalculated whenever the underlying inputs change.
