Why C2PA manifests die on upload
A C2PA manifest is a hash over exact bytes, so re-encodes and screenshots kill it by design. How to find the hop in your pipeline where provenance is lost.
8 min read
Why the editor rejects local file paths
Clip sources must be reachable over HTTPS, so file:// paths fail immediately. Upload or attach first, then render from the hosted URL.
7 min read
Why your clip shimmers, and how to stop it
Shimmer is usually a frame-rate problem: too many frames starve temporal context. Rates that hold, where the deflicker pass goes, and how to spot compression.
9 min read
How to fill a 4,500-token prompt window
Long-window image models take far more instruction than most prompts supply. A block structure for long prompts, and where extra tokens stop mattering.
8 min read
Writing acceptance criteria clients can't argue
Turn 'approved' from a feeling into a test: technical spec, binary brand checks, a bounded subjective layer, a named approver, and a deemed-accepted window.
9 min read
Your cloned voice sounds nothing like you
Clone quality is set by the reference recording, not the model. Source-audio requirements and the three defects that guarantee a bad clone.
9 min read
Your edit model repainted the whole image
Instruction-based edit models regenerate the whole frame. Scope the change, then composite the region back so the background stays put.
9 min read
Every face comes back too symmetrical
Models average toward a symmetrical face that reads as generated. Asymmetry cues to prompt, plus the guidance and checkpoint choices that stop flattening.
8 min read
Your image came back with a watermark you never added
Hallucinated stock marks come from training data, not the platform. Prompt changes that suppress them, plus a clean removal pass when they appear.
9 min read
Your image-to-video clip barely moves
A clip that reads as a photo with drift is an underspecified motion brief. Write one subject action and one camera move, then fix the source frame.
9 min read
Your ducking pumps instead of breathing
Pumping is a release-time problem, not proof sidechain was wrong. Attack, release and depth settings that hide the duck behind the voice.
9 min read
Your prompt was blocked. How to rewrite it
Most prompt blocks come from a few trigger patterns, not the idea. Isolate the phrase, rewrite it, and know when to switch models instead.
9 min read
Your reference image got cropped on upload
Some pipelines hard-crop uploaded references to the target ratio without warning. Pre-pad to that ratio yourself so the crop has nothing to take.
9 min read
The subject teleports halfway through the clip
Position snaps come from independently sampled segments with no shared anchor. First-last-frame conditioning lets motion flow through the join.
9 min read
Your generated subject never blinks
An unblinking subject is the tell that survives every other fix. Schedule eye behaviour in the prompt, or chain short clips at natural blink points.
9 min read
Your TTS mispronounces brand names
Invented and non-English names are the TTS failure case. Phonetic respelling that works across engines, and which providers take pronunciation markup.
9 min read
The upscaler changed your subject's face
A generative upscale redraws identity. Keep face denoise at or below 0.35, mask the face out of a hotter pass, and check a 100 percent crop before you commit.
9 min read
The model added dialogue you never wrote
Native-audio models treat 'no talking' as a cue to talk. Convert every negative into a positive audio state, with worked before-and-after prompts.
9 min read
Your voiceover sounds muddy under AI music
AI music beds crowd the same 250–500Hz band as speech. Carve that range, set a level relationship, and apply them in the right order.
9 min read
What YouTube's AI disclosure does not cover
YouTube exempts animation, beauty filters, colour work, effects, your cloned voice, captions and audio repair. Decide in one pass if a video needs the label.
9 min read
Where YouTube puts your AI label
Since May 2026, photorealistic AI labels sit below the long-form player and overlay Shorts. Unrealistic and minor edits stay in the expanded description.
9 min read
YouTube Shorts length and resolution limits
Shorts cap at three minutes and 1080p. A hard delivery ceiling changes what generation resolution is for, and turns downscaling into a quality decision.
8 min read
A Diagnostic Tree for Generations That Come Back Wrong
A decision tree for AI generations that come back wrong: what to change based on exactly what you see in the output, not what you originally asked for.
7 min read
Why Ad Review Flags Generated Creative (and How to Pass on the First Try)
AI ad creative gets rejected for reasons that trace back to specific platform rules — watermarks, UI mimicry, health claims, missing disclosure. The fixes.
9 min read