Use: “jagged edges, aliasing, pixelated, staircase edges” if your model responds to those terms. The bigger win is preventing low-res outputs and compression artifacts, so keep “lowres, jpeg artifacts” in your negative. Also avoid extreme sharpening cues—those can create edge halos that look like aliasing.

Use stable perspective cues: “eye-level camera, 35mm or 50mm, straight lines, realistic architecture.” Wide-angle close to buildings causes distortion and warping. Keep the camera a bit back, reduce extreme angles, and don’t overload the prompt with too many building details at once.

iPadOS drag-and-drop is inconsistent across browsers and input types. Use the upload button and pick from Files/Photos instead. If you’re dragging from Photos, saving to Files first usually increases compatibility.

Because long prompts often contain hidden conflicts. SDXL tries to satisfy everything, so if you stack multiple styles, lighting setups, camera lenses, and moods, the model averages them into something muddy. A simple prompt creates a clean target. The best workflow is: start small (subject + setting + style), lock a seed, then add only 1–2 details per iteration. If quality drops, rollback the last additions rather than adding even more words.

Eye color is a tiny feature that competes with lighting, makeup, and overall palette. If it’s ignored, you either have too many competing descriptors or the face is too small. Make the face a priority: “close-up portrait, eyes in focus” and keep eye color near the subject identity (“woman with green eyes”). If needed, lightly weight: (green eyes:1.2). Avoid stacking “dramatic lighting” that changes perceived color; “neutral light” helps eye color read correctly.

Use positives like “24mm landscape photo, deep depth of field, everything in focus, f/8”. Avoid positives like “bokeh” or “shallow DOF” because they contradict your goal. If the model still blurs backgrounds, add negatives like “shallow depth of field, bokeh” (only for this preset) and keep the scene description clear and uncluttered.

Then at least use a unique passphrase you’ve never used before and store it offline (written down and kept safe). The dangerous option is “I’ll just reuse something I remember.”

Multiple bodies multiply overlap problems. Keep interactions simple (standing side-by-side), use higher resolution, and avoid hugging or tangled limbs at first. If you need physical contact, generate a clean base, then inpaint the contact region. Multi-person anatomy is one of the hardest tasks; workflow beats prompt length.

Very long prompts can conflict and cause the model to prioritize only some parts. Start with a shorter prompt and add details slowly. Also avoid mixing styles (photoreal + anime) unless you intentionally want a hybrid.

Often caused by extreme angles, hats/head accessories, or too-tight crops. Try a neutral angle first and avoid combining bangs + hat + wind + extreme tilt. A small negative like “distorted hairline, deformed forehead” can help if persistent.

Keep the face close, lighting controlled, then inpaint jewelry if needed. Globally forcing “detailed jewelry” often adds noise everywhere; local repair is cleaner.

False positives can happen. If you believe it’s an error, open a support ticket with the date/time and a brief description of the prompt’s intent (do not include prohibited details).

POSITIVE:\n“photo, adult person, (realistic anatomy:1.12), (natural proportions:1.1), 50mm, eye-level camera, soft key light”\nNEGATIVE:\n“(bad anatomy:1.2), (extra limbs:1.25), (missing limbs:1.2), (deformed body:1.2)”

Use gentle realism cues: “natural hand anatomy, subtle knuckles” and soft lighting. Avoid forcing “hyper detailed knuckles” because it can create harsh bony protrusions. Real knuckles show through directional light and slight skin folds, not through extreme sharpening.

Often the last step is “finalizing” (writing file to disk). Free up storage, close other heavy apps, and retry on a stable connection. ZIP is usually smoother than many separate images.

Fog is subtle. Use “light haze” and don’t stack it with heavy DOF blur.