Simplify composition tags and remove mirror/reflection tags. Add negatives like “extra head, duplicate” and keep seed stable.

Prompt controlled highlights and soft contrast: “soft filmic contrast, controlled highlights, gentle shadows.” Avoid “HDR” and extreme “high contrast” stacks. Lighting words matter: softbox/window light tends to preserve highlight rolloff. If highlights blow out, reduce “glossy highlights” cues and avoid overly shiny materials. Photoreal exposure is about restraint; if you push contrast too hard, it stops feeling like a real camera sensor.

We may email a general policy notice, but we don’t provide the exact matching words or detailed filter rules. That information can be abused to bypass safety.

Say what kind of night: “night portrait, street lit by warm sodium streetlights” or “moonlit outdoor scene, soft low light.” Avoid default cyberpunk triggers like “neon, glowing, futuristic.” Add “low light photography, subtle noise, soft shadows” if you want realism. If you want crisp night detail, increase exposure cues (“soft fill light”) instead of cranking contrast. Realistic night is dark but still readable—give the model a reason the camera can see the subject.

Logs can contain hidden identifiers. GenIDs let staff locate the exact error internally without you sharing potentially private data in a ticket or public chat.

Small metallic objects can confuse anatomy. Start with no jewelry while dialing in hands, then add one piece at a time (ring OR bracelet). Use “simple ring” rather than “intricate ornate ring” at first.

Most autocompletes treat commas as “end of token.” That’s expected—type the next tag after the comma to see suggestions again.

Blurry outputs often come from “soft focus / motion blur” selections or overly aggressive negatives that accidentally block detail. Check if the wizard added blur terms in the positive. Also ensure you’re using an SDXL-friendly resolution; too small can look mushy. If you want sharper, increase resolution and use one “sharp focus” cue, not six.

POSITIVE: RAW smartphone selfie, adult woman, cardigan, plant store, greenery everywhere, soft daylight through windows, arm extended, (sharp eyes:1.2), natural tones NEGATIVE:

Those come from over-smoothing and sharpening halos. Reduce steps/CFG a bit and avoid “perfect skin, flawless” stacks. Add negatives like “beauty filter, skin smoothing, oversmoothed, plastic skin” and use softer lighting words.

Save two prompt presets: one “photo block” (lens/lighting) and one “anime tag block.” Swap the preset rather than trying to edit everything manually each time.

It’s almost always a missing closing brace `}`, missing `endif;`, missing `);`, or a broken heredoc/nowdoc string. It can also be caused by an unterminated comment or quote. The fastest fix is to run the file through `php -l filename.php` (lint) and jump to the line it reports, then check the block nesting above it.

Instant failure is usually input validation (missing `image_id`, missing login session, malformed JSON), policy blocks, or server-selection returning no valid server. Check what your backend returns in JSON: a stable error key like `not_logged_in`, `missing_image_id`, `PROMPT_BANNED`, etc. Make sure you always return a structured JSON error and never die with HTML.

A big one is ordering: earlier tokens often get slightly more attention. Put your identity + camera + lighting early, then environment, then styling. Another is reducing conflicts: if you say “soft light” and also “hard rim light,” you’ll get weird lighting. You can also use very short “anchor phrases” like “portrait photo” or “candid smartphone photo” to push realism, and you can keep a “baseline template” then swap only the variable parts. These aren’t secret hacks—just consistency tools that make results reproducible.

Watches are tiny—give them pixels. Use a closer framing: “close-up wrist shot, watch in sharp focus, studio lighting, controlled reflections.” Don’t expect a detailed watch in a full-body image. If needed, inpaint the watch region like a product photo.

Resolution first. SDXL can’t invent clean micro-detail if the canvas is too small. Raise resolution to an SDXL-friendly size, then adjust steps modestly. For photoreal detail, lighting and focus cues also matter—good light makes real texture visible; bad light forces the model to guess.