Open weights make local image generation possible. Hardware, electricity and maintenance still have a cost.
How Much Does It Cost to Self-Host an Image Model?
Estimate Ming Design and Qwen-Image-2.1 costs by monthly image volume. Choose a measured hardware profile, add your prices and compare local and cloud budgets. License terms vary by model.
Electricity for hardware you already own
This total includes only the electricity for the whole machine during the hours entered. Equipment purchase, depreciation, labor and cloud rental are excluded. It replaces the ownership estimate; do not add the two electricity totals.
Choose the currency on your electricity bill. Changing currency does not convert the rate; enter the new rate yourself.
Enter measured average power for the entire system, including CPU, RAM and PSU losses. GPU TDP is not a whole-machine measurement. No extra PUE multiplier is applied.
Enter elapsed running hours. Divide measured seconds by 3,600. Monthly time is limited to 744 hours; include idle hours only if the same average power covers them.
Use the variable energy rate on your bill. Fixed connection fees are separate.
Enter finite, non-negative power, hours and price. Maximum: 10,000,000 W, 1,000,000 task hours or 744 monthly hours, and 10,000 per kWh. Zero is allowed.
Time is entered directly here. Other workload calculators derive it only where matching performance data exists. Output duration, model memory fit and a GPU name do not establish generation time. For your own setup, measure the complete task, including offload and retries, before estimating its electricity.
Updating estimate…
API workload budget
Estimate metered usage in USD from the stored tariff. Include every attempt and retry. Taxes, credits, free allowances and fixed fees are separate. This is independent of the self-hosting capacity estimate.
Klein generation and editing use the same per-image resolution tariff. Count all attempts, including rerolls. Other models, custom dimensions and separately billed services are not included.
Estimated usage cost: $8.00
1,000 Billable images × $0.008 · 768 × 768
Image models
- Qwen-Image-2.1
- Ming-Image-0.1-Design
- Nano Banana 2.1
- Seedream 5 Flash
- FLUX 3 Image Klein
- MOSS-Image
- Iris-3B and Turbo
- Krea 2 RAW / Turbo
- Krea 3
- Fibo 2.0
- Ideogram Vector 1.0
- GPT Image 2.6
- Pruna P-H3-Image
- Seedream 5.5
- MiniMax H3-Image
Turn your image workload into a monthly budget
See what one image costs at your expected volume.
Upfront hardware spreads across the images you generate. Electricity follows runtime, while maintenance remains part of the monthly bill.
Share the configuration or export the report to discuss the assumptions with your team.
Results show upfront investment, monthly cost, cost per image and cost per 1,000 images.
What determines image generation cost?
Task, resolution, monthly volume and deployment conditions all affect the estimate.
Resolution and steps
Ming profiles cover 1024px and 2048px at 12 steps. Qwen measurements cover 1024px at 40 steps. Each workload uses its own timing sample.
Monthly volume
Images per month × measured seconds per image determines compute time. Fixed costs spread across that volume.
Hardware and offload
Precision, encoder placement and host RAM change the setup. Qwen’s measured 4090 profile offloads encoder layers to CPU; Ming uses separate 3090 quantization profiles.
Power and maintenance
Electricity, cooling, hardware depreciation and operations labor contribute to local running costs.
Adjust exchange rate, electricity, labor, PUE, system overhead and depreciation to match your situation.
The calculator models one deployment. It does not automatically add GPUs when capacity is exceeded.
Method and assumptions
How the workload becomes a cost estimate.
- Monthly compute hours = images × measured seconds per image ÷ 3,600. One image means one generated output.
- Only benchmarks matching task, resolution and steps are used. Throughput does not extrapolate from 1K to 2K or between generation and editing.
- Local monthly cost includes hardware depreciation, active-runtime electricity and operations labor. Cloud estimates use active rental hours when a valid hardware price exists.
- Unit costs divide the monthly total by image volume. Workloads above one deployment’s monthly capacity are marked unavailable.
- Ming Design uses MIT. Qwen-Image-2.1’s research license covers non-commercial research or evaluation; its separate commercial license fee is excluded.
Costs to budget separately
Budget separately for costs outside the measured inference workload.
- For Qwen commercial use, budget separately for the required license agreement.
- Image quality and editing consistency are not scored by the calculator.
- Moderation and compliance work are not included.
- Storage, delivery bandwidth and asset management add costs.
- Failures, retries, cold starts and idle rental time are not modeled.
- Prompt length, reference images, runtime and concurrency can change generation speed.
Estimate your image generation budget
Start with a measured workload, then adjust the costs you know.
- 01
Choose a task and volume
Select generation or editing, output resolution, steps and images per month.
- 02
Set hardware and assumptions
Pick a documented hardware profile and enter your electricity and maintenance costs.
- 03
Review and share
Review monthly and per-image costs and license terms, then save the configuration or export the report.
Self-hosting or a hosted API?
Cost is one part of the choice. Control, operations and licensing also matter.
Self-hosting may fit when:
- Monthly volume is steady enough to use the hardware.
- You need control over input data and the inference pipeline.
- The model meets your image quality requirements.
- Your team can operate GPU inference.
- The exact checkpoint and components are licensed for your intended use.
A hosted API may fit when:
- You want to avoid maintaining GPU servers.
- Image volume changes substantially each month.
- Provider output quality and features match your workflow.
- The exact model, billing terms and data policy are documented.
- The provider terms cover your intended use.
Supported image models
Compare model components and read the license terms.
Data sources
Model facts come from publisher documentation. Timing profiles cite their original experiment and hardware setup.
Recorded prices include source, region, date and billing conditions. Quotes you enter are labeled separately.
Qwen profiles checked September 21, 2026; Ming profiles checked October 1, 2026.
FAQ
About Ming Design, Qwen-Image-2.1 and the measured cost profiles.