Every AI generation is a dice roll. Same reference, same prompt, same settings — and the four outputs still differ: a hand lands better in one, the light falls better in another, the expression is right in a third. Batch generation is the feature that turns that randomness from an enemy into a tool: instead of rolling once and re-rolling on failure, you roll four or eight dice at once and keep the best. Here is why that beats prompt perfectionism, and the workflow that gets the most out of it.
The math nobody does
Suppose a decent prompt gives you an image you would actually keep 40% of the time. Generate once, and you are more likely than not to be re-prompting. Generate a batch of four, and the chance that at least one output is a keeper is about 87%. At batch eight it is 98%. Nothing about your prompt improved — you just stopped betting everything on one roll.
Now compare the alternative path: spending twenty minutes polishing the prompt to push that 40% to 55%. A single roll at 55% still loses to four rolls at 40%. Past "good prompt", refinement buys less than repetition does. This is why experienced users obsess over selection, not perfection.
What a batch tells you that a single roll can't
Beyond raw hit rate, a batch is a diagnostic. When one output has a warped hand, you learn nothing — flukes happen. When all four do, the problem is systematic: the pose you asked for is hard, or your negative prompt needs an entry. A single generation cannot distinguish bad luck from a bad input; a batch of four does it instantly.
The same logic applies in reverse to your prompt experiments. Changed a wording and the result improved? On one roll, that is noise. On four, it is signal.
The workflow
A batch workflow that doesn't waste credits
- 1
Draft at batch 1
While choosing your reference, pose, and scene, generate singles. No reason to pay 4× to discover you want a different location.
- 2
Lock the scene, switch to batch 4
When a single roll shows the right composition, keep every setting and re-run as a batch. You are no longer exploring — you are sampling.
- 3
Select on face first
Compare the batch at full size: face, then hands, then background. Viewers forgive a soft background; they never forgive a broken face.
- 4
Save the winner, note the pattern
If one output wins for a repeatable reason ('three-quarter angle works for this character'), it feeds your next draft.
- 5
Escalate to batch 8 only for hard scenes
Complex poses and busy lighting have lower hit rates — that is where the widest net pays for itself.
The honest cost math
A batch of four costs four generations — Gensomnia does not discount volume, and pretending otherwise would be marketing. The saving is not in credits; it is in everything around them: one wait instead of four, one decision point instead of four re-prompt cycles, and fewer total generations spent, because selection converges faster than iteration. Users who re-roll singles until something lands routinely spend more credits than a drafted-then-batched flow, and always spend more time.
Batch size sits in the generation settings panel (the button next to Generate) alongside aspect ratio and the negative prompt — the Pro toolkit. The panel is visible to every user, so you can see the controls before deciding whether they are worth it.
If you are still setting up your basic flow, start with the 30-second quickstart and a solid prompt structure — batching multiplies a good setup, it does not rescue a bad one.
Frequently asked questions
What is batch generation in AI art?
Generating several images from the same inputs simultaneously. Because outputs vary naturally, a batch gives you multiple candidates to choose from in one run.
Does batch generation cost more?
Yes — a batch of four costs four generations. The savings are in time and total attempts: selecting from variants converges faster than re-prompting singles.
What batch size should I use?
Draft at 1 while exploring, switch to 4 once the composition is right, and reserve 8 for difficult scenes with low hit rates.
Why do identical prompts produce different images?
Generation starts from randomness by design; the prompt and reference constrain the outcome but never fully determine it. Batching turns that variance into choice.
Is it better to improve my prompt or generate more variants?
Both, in order: get the prompt to 'good' first, then let batches do the last mile. Past a decent prompt, selection improves results faster than further wordsmithing.
Where is batch size in Gensomnia?
In the generation settings panel next to the Generate button, with steps of 1, 2, 4, and 8. It is part of the Pro settings, visible to all users.