iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Generating thousands of images did not finish Whispart Studio’s artwork; choosing among them did. In its account of the process, the studio describes reviewing model checkpoints, generating a much larger candidate batch from a promising state, and narrowing that batch through human selection. The counts are approximate recollections, not audited totals for the featured work.
Why choosing mattered more than generating
A generative model can produce a large set of plausible images, but plausibility does not decide which image remains compelling after repeated viewing. Whispart frames the central question as: “What do you use to tell a generative system’s most interesting outputs from its most immediately polished ones?” Its answer is a human-led process of comparing, revisiting, and selecting—not a claim that the system automatically identifies the best result.
The studio’s account separates two decisions: first, which point in training is worth exploring; then, which individual image from that point is worth keeping. Technical progress can inform the first decision, but it does not settle the second.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsStage one: review checkpoints during training
As described in the Whispart article, the team looked at sample grids saved at training checkpoints. The founder recalled saving a model state at roughly 100-kimg intervals. The article explains “kimg” as thousands of real images shown to the discriminator during training; the interval is an approximate studio recollection, not a universal GAN setting.
#1 Best Overall
- Used Book in Good Condition
To compare states, the studio used about 50–100 examples generated from fixed seeds. Keeping seeds fixed makes it easier to see how outputs change between model states, rather than confusing changes in the model with changes caused by sampling different seeds. This review helped the team notice visual developments and decide whether a checkpoint merited closer exploration.
A checkpoint was not treated as a score that must improve with more training. In the article’s words: “For us it means the checkpoint is a choice, not a score that always goes up.” A later state might be technically different without being more interesting for a particular artwork.
Rank #2
Stage two: select from a much larger batch
After settling on a promising state, the studio account says it might generate on the order of 10,000 candidate images. The article describes two human reduction rounds: two teammates first narrowed the batch to roughly 1,000, then the founder selected roughly 100.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →| Stage | Approximate scale reported | Purpose |
|---|---|---|
| Checkpoint review | About 50–100 fixed-seed examples; checkpoints recalled at roughly 100-kimg intervals | Compare visual changes across model states and decide which state to explore. |
| Candidate generation | On the order of 10,000 images | Build a large pool from a promising model state. |
| First reduction | Roughly 1,000 images | Two teammates narrow the candidate pool. |
| Final selection | Roughly 100 images | The founder makes a further selection. |
All figures in the table are approximate recollections reported in the Whispart article, not audited counts or a recommended formula. The account does not establish that these exact quantities applied to the featured work, or that another studio should use the same batch sizes.
What the selected image asks the viewer to notice
Whispart names “Unnamed Heir” as one selected work. The article describes a pale figure that reads quickly against a dark ground, while shifting edges take longer to resolve. The title is offered as an opening for interpretation, not as a full explanation of the image.
The example illustrates why a fast first impression may not be the only selection criterion: an image can reward a longer look. It does not prove which checkpoint produced it or verify the candidate counts described in the workflow.
What a selection interface needs to preserve
The practical interface advice in the studio account follows from comparing large numbers of similar outputs. Reviewers need to be able to revisit an image, understand where it came from, and avoid losing candidates they have not decided about yet.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- Keep checkpoint and seed identities attached to each image so a selection can be traced back to its source.
- Make full compositions easy to open from a grid; thumbnails alone may conceal details that matter during a closer comparison.
- Provide a “hold” state for undecided images, rather than forcing an immediate keep-or-discard choice.
- Leave the final selection to people. The interface can support comparison and organization, but the account presents aesthetic judgment as human work.
These are workflow aims described by Whispart, not a comparison of software products or a tested standard for GAN review.
Best Value
What the account does—and does not—establish
The article’s value is its distinction between monitoring training and selecting finished candidates. It gives a concrete account of how one studio approached those stages, while presenting its figures as approximate. It does not demonstrate that more training always improves an image, establish an ideal number of candidates, or provide an independently audited history for “Unnamed Heir.”
Read the full studio account at Whispart’s article on DEV Community, published September 24, 2025.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
