How do I choose a rented GPU for a batch of ComfyUI images?

Start with three prompts from the real batch: one typical image, the hardest composition and the highest delivery resolution. Run them on the same checkpoint and graph. Record peak VRAM, wall time, accepted outputs and how often you need another seed. A card that costs more per hour can cost less per accepted image if it finishes the same batch sooner.

What changes the card requirement?

Workload First comparison Pilot measurement
SDXL drafts at a fixed resolution A 24 GB card running a single image at a time Warm render time and rejection rate
FLUX with extra LoRAs, ControlNet or a second pass Compare 24 GB with a larger-memory configuration supported by the graph Peak VRAM at the heaviest stage and CPU offload time
Many independent prompts Run the same manifest serially before trying multiple workers Startup overhead per worker and accepted images per billed hour

Keep seed, resolution, sampler, step count and model revision with each output. ComfyUI queue length and model batch size are separate controls: queue separate jobs first when VRAM is tight. FLUX and SDXL memory guide · GPU rental options

Price the batch you can use

Suppose a brief needs 40 approved images. If the pilot produces 4 acceptable results in 5 attempts, a planning estimate is 50 attempts. Multiply by measured warm time, add preparation and export, then apply the terms of the live rental quote. Count separately priced storage once. Recalculate after the first ten production images if the acceptance rate changes.

Use stable filenames such as sku-view-seed.png and keep a manifest of approved files. Open the exported images at delivery resolution before release. For catalogue work, follow the product photo masking guide.