Calabi Labs · Guide · 2026-06-19
Seedance 2 0 prompts
Seedance 2.0 is Claude's latest AI video generation model, and the prompts you write for it determine whether you get a polished, usable clip or a blurry mess. The model responds to specific descriptive language, structured scene composition, and motion cues — here's exactly what works and why your output might be getting flagged once you post it.
What Seedance 2.0 Actually Is
Seedance 2.0 is an AI video synthesis model that generates short video clips from text descriptions. It sits alongside models like OpenAI's Sora, Kling, and Runway in the creative-video space. The 2.0 iteration improved temporal consistency — meaning motion across frames holds together better — and handles longer clip generation than its predecessor. Creators use it for B-roll, social content, product visuals, and narrative shorts.
The quality of output depends almost entirely on how you prompt it. A vague prompt gets you generic output. A precise, structured prompt unlocks the model's full capabilities.
What Actually Gets Flagged When You Post AI Video
Before diving into prompt writing, understand the downstream problem: even a perfectly crafted Seedance 2.0 video can get flagged, shadowbanned, or suppressed on Instagram, TikTok, YouTube, or Reddit. That happens at the metadata and signal layer — not the visual layer.
Platforms scan for three invisible things in your uploaded file:
- C2PA / Content Credentials — A cryptographic manifest embedded in the file (stored as JUMBF atoms) that explicitly declares "this was generated by AI." Seedance exports carry this by default.
- XMP AI flags — Tags like
DigitalSourceType: trainedAlgorithmicMediaembedded in the file's metadata, which most platforms check automatically. - Encoder fingerprints — Seedance exports through specific encoders (like Lavc or x264 with SEI NAL units) that leave recognizable fingerprints in the bitstream itself, signaling machine generation even when metadata is stripped.
A raw Seedance 2.0 export carries all three. Platform scanners detect these signals within seconds of upload, before any human ever sees your content.
Why Standard Edits Don't Fix the Flagging Problem
Creators try three common workarounds — none of them solve the actual issue:
- Cropping the video — Removes a visible corner watermark if present, but leaves the C2PA manifest and XMP flags completely intact in the file structure.
- Screenshotting or re-recording — A screen recording still carries the encoder fingerprint and can still embed C2PA if the player injects it during playback.
- Re-exporting through editing software — Tools like Premiere or DaVinci Resolve strip some metadata but don't touch C2PA manifests embedded at the codec level, and they don't remove encoder fingerprints.
The invisible signals survive all of these. That's why your Seedance export can be flagged even after you've cropped it and run it through editing software.
How to Write Seedance 2.0 Prompts That Actually Work
Structure your prompt in layers: subject, environment, motion, camera, and style. Each layer gives the model more to work with.
1. Subject and Action
Start with who or what is moving. Be specific about the subject's characteristics rather than generic. "A woman in a red jacket" performs better than "a person." Specify the action precisely: "walking slowly through rain" beats "walking outside."
2. Environment and Lighting
Describe the setting and light source. "Golden hour light streaming through a dusty warehouse window, volumetric rays visible in the air" gives the model strong visual direction. Specify time of day, weather, and key light position for consistent results.
3>Motion and Dynamics
Seedance 2.0 handles complex motion better than earlier versions. Describe movement vectors explicitly: "the camera pushes in slowly while the subject turns to face it" or "water droplets bead and slide across a dark surface, individual drops catching highlights." Reference physics — gravity, momentum, fabric dynamics — for realism.
4. Camera and Framing
Use standard cinematographic language. "Medium shot, shallow depth of field, bokeh background" or "low angle, wide lens, tracking shot following the subject." Seedance 2.0 understands terms like "dolly zoom," "crane shot," and "single-point perspective."
5. Style and Aesthetic
Reference visual references if needed: "filmic grain, inspired by Wong Kar-wai cinematography, warm teal and orange grading." For hyperrealistic output, specify "photorealistic, 8K, cinematic color science."
Example Prompt Structure
A woman in her thirties with long dark hair, wearing a weathered leather jacket, walks slowly down a rain-slicked Tokyo backstreet at night. Neon signs reflect in the puddles. She pauses under a buzzing yellow light and looks up. Camera: slow push-in, shallow depth of field, bokeh on the background signs. Style: cinematic, moody, inspired by Blade Runner 2049, hyperrealistic lighting, 8K render.
After the Generate — Cleaning Your Seedance 2.0 Export
Once you have a Seedance 2.0 clip you're happy with, run it through Calabi before posting. Calabi's pipeline strips the C2PA manifest, removes the DigitalSourceType: trainedAlgorithmicMedia XMP flag, and injects authentic phone-capture identity — GPS coordinates, capture timestamp, a real device profile like iPhone 16 Pro or Pixel 8 Pro, and a real encoder name.
You get back a forensic proof card showing exactly what was stripped and what was injected, so you can verify the clean status before uploading. A raw Seedance 2.0 export has 18 JUMBF atoms and 144 metadata tags. After Calabi, it sits at 0 JUMBF atoms and about 94 neutral structural tags — the same profile as a real phone recording.
FAQ
Does Seedance 2.0 embed C2PA by default?
Yes. Seedance 2.0 exports include C2PA / Content Credentials metadata declaring the file as AI-generated. This is embedded at the codec level and survives cropping, screenshotting, and most re-encoding workflows.
Can I remove the AI detection signals manually?
Stripping C2PA manifests and XMP AI flags requires a tool that parses and rewrites the file at the metadata layer. Standard editing software doesn't touch C2PA. Calabi handles this automatically and also replaces encoder fingerprints that metadata-only tools miss.
Will cropping remove all AI signals?
No. Cropping removes visible content and can remove a visible corner watermark, but the C2PA manifest and XMP flags survive because they're stored in the file structure, not the pixel area. The encoder fingerprint also remains regardless of frame cropping.
Try Calabi free at calabilabs.com — 10 cleans, no card.