Remove “hair_over_one_eye” type tags and add “clear eyes, eyes visible” (sentence). Use a hairstyle that keeps the face open (side-swept bangs minimal).

Because sharing exact triggers helps bad actors evade detection. It’s safer to show high-level categories (policy violation, prohibited prompt) rather than the exact matching logic. If you believe a restriction was a mistake, use the support ticket system with your username and relevant details. That gives moderators what they need without exposing the filter rules publicly.

Download ZIP. One file avoids multi-download blocking, reduces missing files, and makes it easier to keep sets together. Use individual downloads only when you want a few selects.

Lock a baseline preset: same checkpoint, same resolution/aspect, same sampler/scheduler, same CFG range, same camera block, and fixed seed when iterating. Pony models can swing style easily; consistency comes from reducing degrees of freedom and changing one variable at a time.

A database error happens when the site can’t read or write required data at that moment (tokens, queue row, saved image record, etc.). First try again after a short wait. If it happens repeatedly, submit a ticket with the time it happened so we can match it to server logs.

That usually means the generation or post-processing failed server-side. Try again with simpler settings (lower steps/resolution) or switch servers. If it repeats, open a support ticket and include the approximate time and your selected mode.

Long prompts create contradictions and dilute importance. The model tries to satisfy everything and ends up satisfying nothing well. Put the most important concepts early and cut anything you’re not sure matters.

Negatives suppress common failure modes (extra fingers, fused fingers), but they don’t tell the model what a correct hand pose should look like. Positive hand phrasing (“relaxed hands, fingers slightly separated, natural hand pose”) provides a target, which is especially useful when the pose is ambiguous. In practice, you use both: a tiny negative block to reduce typical errors, and a simple positive pose description to guide structure—while keeping the scene uncluttered to preserve hand edge clarity.

GenID is a unique ID attached to a single request/error. It lets support pull the exact server-side logs for your run (without you needing to explain everything). When you contact support, paste the GenID plus a quick summary like “img2img failed” or “queue not found.”

Use “35mm group photo, f/4 or f/5.6, everyone in focus, level horizon”. Avoid “shallow DOF”. Keep lighting simple and specify the number of people. If extra people appear, add a small negative “extra person” and keep the environment simple. Group shots are mostly composition + DOF; lens tags help a lot.

Not really. Tags describe what you want; negatives help block what you don’t want (artifacts, bad anatomy, text). They work together.

Add negatives for “heterochromia” if unwanted, and keep face straight-on.

Use: “solo, portrait, looking_at_viewer” then add “DSLR photo, 85mm, soft window light, natural skin texture, shallow depth of field.” Keep tags minimal.

High CFG can force unnatural detail and distort anatomy. For photoreal, CFG 4.5–7 is usually safer.

It pushes a tighter crop on face/shoulders. Great for beginners because it reduces anatomy complexity.

Don’t click suspicious links. Open the website manually and log in normally. If there’s an issue, you will see it in your dashboard or via support tickets.