The model had finished generating. That did not mean we had a finished artwork. In Whispart's studio workflow, it meant we had a much more difficult question: which image would still be interesting after the first two seconds?
We make images with a GAN model we call Mnemosyne. Our founder describes a working pattern with two very different kinds of looking. During training, the team reviews sample grids at saved checkpoints; after choosing a promising model state, it explores a much larger batch of outputs. Those numbers are approximate studio recollections, not a public benchmark or a recipe that every GAN project should copy.
The first question: when does more training stop helping the art?
In the founder's account, a model state may be saved at roughly 100-kimg intervals. Here, kimg measures thousands of real images shown to the discriminator during training (NVIDIA's StyleGAN2-ADA documentation). The studio uses about 50–100 fixed-seed examples to watch how the visual field changes. Fixed seeds make a comparison across checkpoints more meaningful: you can see what changes in the same starting positions rather than confusing a different sample with a different model.
But the comparison is aesthetic as well as technical. A later state can be more resolved while losing the surprising, half-formed relationships we value. That does not mean less training is universally better, or that a specific kimg value guarantees good art. For us it means the checkpoint is a choice, not a score that always goes up.
The second question: what survives a large batch?
Once a state looks promising, the studio may generate on the order of 10,000 candidate images. In the founder's description, two teammates first reduce that field to roughly 1,000 images that feel distinctive, compositionally alive, or worth imagining on a wall. The founder then makes a second selection of roughly 100. These are working estimates, not audited counts for the individual image below.
The first pass asks, “Is there something here?” The next asks, “Does it keep happening when I look again?” A clean thumbnail can fail the second question. A visually unstable image can pass because its composition leaves room for discovery.
Unnamed Heir is one selected Whispart work. The pale figure reads quickly. The dark ground and shifting edges take longer. The title offers a narrative opening without telling you whose heir this is or what she will do next. We are showing it as an example of the kind of image that rewards a second look, not as proof of an exact checkpoint or candidate count.
What we would want a review tool to preserve
If you build generative tools, here is the practical lesson from our process: the interface should help a human compare and return, not simply reward the output that wins the first glance. Keep checkpoint and seed identities attached to images. Make the full composition easy to open from a grid. Give “hold” its own state so an undecided image is not silently treated as a rejection.
And keep the final decision separate from generation. A model can produce a compelling possibility. People still decide what to keep, name, test in material form, and release. The Whispart process journal shows how those stages fit together.
What do you use to tell a generative system's most interesting outputs from its most immediately polished ones?
Published by Whispart. This article was drafted with an AI writing assistant using the founder's account and the studio's public process material, then checked against those sources. The figures above are approximate descriptions of a working pattern, not a claim about the provenance of the pictured artwork.
