image/flux-schnell@1

One prompt in, one PNG out, from black-forest-labs/FLUX.1-schnell with its transformer and T5 encoder quantized to 4-bit NF4 (computing in bfloat16), without guidance (the model is distilled to run without it).

A unit

One image at 1024x1024 in 4 steps (the default, which is also the largest unit), or smaller: width and height are multiples of 16 from 512 to 1024, and steps run from 1 to 4. A unit outside these fails with the reason before the GPU does any work.

Manifest fields

Field	Required	Meaning
prompt	yes	the text to draw; the CLIP encoder reads its first 77 tokens and the T5 encoder its first 256
seed	no	an integer that fixes the starting noise; 0 by default
width, height	no	1024 by default
steps	no	4 by default
unit_id	no	your name for the unit: 1 to 64 letters, digits, ., _ or -, starting with a letter or digit; its results are stored under it. By default units are numbered by line: 000000, 000001, …

Example

prompts.jsonl:

{"prompt": "a macro photo of a watch movement", "seed": 4}
{"prompt": "a lighthouse in a storm, oil painting", "seed": 5, "width": 768, "height": 1024}

job.yaml:

name: posters
template: image/flux-schnell@1
input: prompts.jsonl

What comes back

image.png for each unit, at the size asked for. Each unit’s directory also holds done.json: {"unit_id": ..., "artifacts": ["image.png"], "seconds": ...}.

Limits

• A GPU with at least 24 GB of memory, 4 CPUs, 16 GB of memory and 40 GB of scratch disk.
• 180 s per unit. A unit still running then is stopped and tried again, up to the job’s attempts (unit.max_attempts: 3 by default, 1 to 10), and a unit that never completes is not billed.
• Checkpointing is allowed, but it is best effort today: interrupted units usually start again from the beginning (see guarantees). A unit is billed once either way.
• A manifest line is at most 65,536 characters and a manifest at most 32 MiB.

Price per unit: see pricing.

Benchmarks

The price is measured on default units (1024x1024 in 4 steps, four prompts), on each GPU class in the fleet with 24 GB or more. The timings are measured on GPU hosts at launch.