Calabi Labs · Guide · 2026-06-19
Crafting prompts for ai generated images
Writing effective AI image prompts is part craft, part vocabulary — you describe the visual you want clearly, use the right modifiers, and let the model do the heavy lifting. But even the most perfectly crafted prompt produces a file that carries invisible "made by AI" signals, and that's where most creators hit a wall when posting online. Here's what prompt engineering actually involves, and what happens to your file after you hit generate.
What Goes Into a Strong AI Image Prompt
A good prompt has three layers. The subject — a person, object, scene — is the anchor. The style — cinematic still, watercolor, editorial photograph — tells the model the visual register. The modifiers — lighting, camera angle, film grain, depth of field — give you control over the mood and technical qualities. Put them together as a natural sentence, not a keyword salad: "A lone figure walks through fog on a coastal road at dusk, shot on 35mm Kodak Portra, cinematic lighting, shallow depth of field."
Most modern models — Midjourney v7, DALL-E 3, Stable Diffusion XL, Flux — handle natural language well. You don't need to game the system with nonsense syntax anymore. What matters is being specific about visual details and consistent with your style across a series. A cohesive body of work reads differently to an audience than a random collection, even when each image was generated identically.
What Actually Gets Your File Flagged After Generation
Here's the part most prompt guides skip: what happens after you download your image. Every AI generation pipeline — Midjourney's, OpenAI's, Stability AI's — writes metadata into the output file. This metadata is what platforms like Instagram, TikTok, YouTube, and Reddit scan for, often within seconds of upload.
The signals that trigger detection fall into several categories. C2PA / Content Credentials — stored as JUMBF atoms in the file — are cryptographic manifests that say "this image was AI-generated" and often name the exact model. XMP metadata fields like DigitalSourceType: trainedAlgorithmicMedia are direct flags. Encoder fingerprints — Lavc, x264 SEI messages in video; specific quantization tables in images — differ between a real phone capture and an AI export. And files lack the GPS + capture timestamp combination that every phone photo carries, which is itself a negative signal.
A Midjourney export might carry 144 metadata tags. A real iPhone photo might carry 20. That gap is what automated detection systems look for — not just what was added, but what is conspicuously missing.
Why Cropping, Screenshots, and Re-Uploading Don't Fix It
Creators often try the obvious workarounds. Crop the image and the visible watermark disappears — but the metadata survives intact. Screenshot the AI image with your phone and you get a fresh phone photo with new metadata — but the encoder fingerprint of the screenshot, the resizing, and the lack of original capture context still read as suspicious. Re-upload to a second platform and download again — you've added another layer of compression but not removed the C2PA manifest or XMP flags baked into the file structure itself.
These approaches change what the image looks like but leave the invisible detection layer untouched. Platform scanners don't look at pixels the way a human does — they read metadata, check for Content Credentials, and evaluate encoder fingerprints. No amount of visual re-encoding addresses that layer.
How to Actually Clean an AI-Generated Image Before Posting
Calabi runs a one-pass pipeline that handles all three stages of forensic cleaning:
- Strip — Remove C2PA / Content Credentials (JUMBF atoms), XMP AI flags (
DigitalSourceType), generator tool tags, and encoder fingerprints that identify AI pipelines. - Inject — Write authentic phone-capture identity: device make/model (iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra), software version, real GPS coordinates, capture timestamp, and a genuine-phone encoder name.
- Verify — Return a forensic proof card — the same ExifTool scan that newsrooms and platforms use — showing exactly what was stripped (18 JUMBF atoms to 0, 16 C2PA references to 0,
trainedAlgorithmicMediaremoved) and what was injected, before you download.
The cleaned file reads, at the metadata level, exactly like a photo your phone took. The C2PA manifest is gone. The XMP flags are gone. The encoder fingerprint matches a real device. A platform scanning that file sees a normal phone photo — not an AI export with 144 metadata tags and a cryptographic "made by AI" manifest buried inside.
Frequently Asked Questions
Does cleaning my image guarantee a platform won't flag it?
No tool can make that guarantee — results vary by platform and by the source model. Calabi removes the metadata and encoder signals that automated scanners specifically look for. Visible watermarks (a logo, a sparkle icon) still need to be cropped out separately, since Calabi doesn't edit pixels. What Calabi eliminates is the invisible detection layer that survives cropping.
Does re-encoding the image through Photoshop or Figma remove AI metadata?
Exporting through a photo editor removes some metadata but almost never strips C2PA / Content Credentials or DigitalSourceType flags — those are stored in specific metadata namespaces that standard export dialogs don't touch. You'd need a forensic-grade stripper to reliably remove them, and even then you'd still be left without phone-capture identity (GPS, timestamp, device profile) that the cleaned file needs.
Do I need to change my prompt or generation settings?
No. Your prompt craft stays the same — that's entirely between you and the model. Calabi operates on the output file after generation. The cleaner your prompt, the better your image; Calabi handles the metadata layer so the file posts cleanly.
Try Calabi free at calabilabs.com — 10 cleans, no card.