Unit economics
Self-Hosting Flux and Qwen vs Paying Per Image: The Real Unit Economics
Flux and Qwen Image Edit both have open weights, so every seller with a growing catalogue eventually asks: should we run them ourselves? The honest answer depends on three numbers you can estimate before buying anything — GPU cost per hour, images per hour, and how busy the GPU actually is.
The formula
Cost per image = GPU cost per hour ÷ (images per hour × utilisation), plus engineering time, storage and failed generations.
Worked example with stated assumptions (not a benchmark): a rented high-end GPU at $2.50 per hour, 10 seconds per edit, gives 360 images per hour. At 100% utilisation that is about $0.007 per image. At 25% utilisation — realistic for a team that edits in bursts — it is about $0.028 per image. Compare that with the live pay-per-image list prices below.
| 模型 | 廠商 | 每張 | 最適合的電商任務 |
|---|---|---|---|
| Qwen-Image Edit Plus | Qwen | $0.032 | 標籤／包裝文字修改、局部重繪 |
| Qwen-Image Edit | Qwen | $0.048 | 懂文字的編輯、去除雜物 |
| Qwen Image 3.0 Edit | Qwen | $0.045 | 多行文字、中日韓包裝 |
| Flux Kontext Dev | Black Forest Labs | $0.038 | 鎖定商品換背景、重打光 |
| Flux Kontext Dev Lora | Black Forest Labs | $0.045 | 品牌風格 LoRA 場景 |
| FLUX.2 Pro Edit | Black Forest Labs | $0.045 | 主圖、寫實生活情境 |
| FLUX.2 Flex Edit | Black Forest Labs | $0.075 | 精細控制、字體排版 |
| Seedream v4 Edit | ByteDance | $0.041 | 批量變體、套圖 |
| Seedream v4.5 Edit | ByteDance | $0.054 | 高解析套圖、一致性 |
| Tencent Image Upscaler | Community | $0.036 | 平台放大檢視（2000px 以上） |
| Youchuan V8.2 Remove Background | Midjourney | $0.13 | 一鍵去背 |
本站 65 個按張計費的修圖與圖片工具模型，定價從每張 $0.0060 到 $0.22。沒有月費訂閱，按產出張數付費。
The costs people forget
- Idle time. A GPU billed per hour costs the same while nobody is editing. Utilisation, not raw speed, decides the bill.
- Engineering. Model updates, queueing, upscaler and cutout models, monitoring. A few days of engineering per month outweigh years of API fees for a small catalogue.
- Licences. Qwen-Image weights are released under Apache 2.0. FLUX.1 Kontext [dev] weights ship under a non-commercial licence; commercial self-hosting needs a licence from Black Forest Labs. Hosted API calls are covered by the provider. Read the current licence text before deploying.
- Retries. Your effective cost is per accepted image. A model with a 50% accept rate doubles every figure above.
When self-hosting wins
- Steady, high volume — tens of thousands of edits a month, every month.
- You need a custom fine-tune or LoRA on your brand style (see LoRA training) and want it on dedicated hardware.
- Data-residency rules require images to stay on your infrastructure.
When pay-per-image wins
- Bursty work: launches, seasonal refreshes, new suppliers.
- You want several models (Flux for scenes, Qwen for text, an upscaler, a cutout model) without operating each one.
- The catalogue is under a few thousand SKUs.
At current list prices, editing a full eight-image set for 100 SKUs is about $25.20 on Qwen Image Edit Plus. If that number is small next to an engineer's week, pay per image. If you need dedicated capacity, Supercomputer runs batches on reserved GPUs.
常見問題
Is it cheaper to self-host Flux or Qwen?
Only at high, steady utilisation. At low utilisation the idle GPU hours make self-hosting more expensive than paying per image.
Can I use Flux Kontext dev commercially if I host it myself?
The FLUX.1 Kontext [dev] weights are under a non-commercial licence; commercial self-hosting requires a licence from Black Forest Labs. Check the current terms.
What licence is Qwen Image Edit under?
Qwen-Image and Qwen-Image-Edit weights were released under Apache 2.0. Confirm the licence of the exact version you deploy.
What is the per-image price if I don't self-host?
On this site, list prices for edit and image-tool models run from $0.0060 to $0.22 per image, with no subscription.