How is total time estimated?

The planner estimates seconds per image from the model, GPU, resolution, and step count, then multiplies by your image count. It is a planning estimate — real runs vary with sampler, VRAM headroom, and whether you batch images per call.

When is cloud cheaper than local?

Cloud wins for occasional large bursts where you do not own a fast GPU, or when you need many images quickly and can rent a high-end card by the hour. If you generate constantly, owning a 3090/4090 amortises fast and local becomes far cheaper per image.

Does a higher step count always look better?

No. Modern samplers plateau around 20–30 steps for most models; beyond that you mostly burn time and credits. Drop steps first when you need to cut a large batch's cost.

How do I cut the cost of a big batch?

Lower resolution and upscale the keepers afterwards, reduce steps to the sampler's sweet spot, generate a small test batch to fix the prompt before committing, and batch multiple images per GPU call to amortise model load.

Batch Image Generation Planner

Batch image generation planning

Generating one image is cheap; generating two thousand for a product catalog, game assets, or a dataset is a real time-and-money decision. This planner estimates the total time, the cloud credit cost, and whether local or cloud hardware is the smarter call for your batch.

How it works

The core estimate chains three numbers:

sec_per_image  = base(model, gpu) × res_scale × step_scale
total_time     = sec_per_image × image_count
cloud_cost     = (total_time / 3600) × gpu_hourly_rate

Resolution scales with pixel count and step count scales linearly, mirroring how diffusion actually works (each step is one UNet pass). For cloud, the planner applies a representative hourly rate for the GPU tier; for local, the marginal electricity cost is small, so the recommendation hinges on whether you already own the card.

Tips for cheaper, faster batches

Prototype small. Lock your prompt and seed range on 4–8 images before launching thousands.
Right-size resolution. Generate at the model’s native res and upscale only the keepers — cheaper than generating everything large.
Trim steps to the sweet spot. 20–30 steps is plenty for most samplers; extra steps mostly cost money.
Batch per call. Generating 4–8 images per GPU call amortises model load and raises throughput on cloud instances.