Rephrase the flagged phrase
“Barefoot” instead of “bare feet.” “Informal attire” instead of body descriptions. “Elderly man” instead of specific anatomical terms. Swap the word, not the meaning.
Safety Blocks
Every AI image, video, and voice model runs its own content-safety filter. When a filter decides a prompt or its output contains flagged material, the model refuses to return it — a safety block. PrePrompt surfaces this on the node so you know why the generation stopped.
Safety filters are not about the user or the creative intent. They’re about specific words and patterns the provider has trained into its filter. A filmmaker working on a mature theme can easily trip one without trying. This page covers how to recognise a block, rephrase around it, and keep moving.
What a safety block looks like
Section titled “What a safety block looks like”When a model blocks a generation, the node shows a distinct status — usually a warning or blocked indicator, separate from a regular failure. The node’s details panel includes the reason the model gave back.
Common reasons you’ll see:
- Content policy — the prompt or output contains material the filter flagged
- Unsafe image / video — the output was generated but flagged before return
- Safety — generic block reason, no specific category given
Unlike a failed generation (which you can retry unchanged), a safety block will block again on retry if you don’t adjust the prompt. The filter is deterministic — the same input will produce the same block.
Why a block happens
Section titled “Why a block happens”A few common patterns trigger filters more often than you’d expect:
- Physical descriptors — “bare skin,” “shirtless,” “exposed” in any context
- Contact words — “touching,” “embracing,” “holding” depending on surrounding context
- Setting combinations — bedroom plus character description can trip filters even for innocent scenes
- Ambiguous pronouns around children — any description involving minors is filtered aggressively
- Weapons in close-up — blades, firearms, blood
- Named public figures — many models block any rendering of specific named people
- Violence or injury — even stylised or historical
- Brand names and trademarks — logos, products, characters from other franchises
None of these are off-limits thematically — they’re word-level triggers. Rephrasing almost always works.
How to reshot around a block
Section titled “How to reshot around a block”Tighten the description
“Two characters seated across a table” leaves less room for a filter to misread than “a man and a woman alone together.” Specificity disambiguates.
Remove surrounding ambiguity
If the subject is adult, say so: “a woman in her thirties.” If the setting is neutral, name it: “living room with visible windows.” Filters err on the safe side when context is missing.
Break complex scenes apart
A scene with four flagged elements won’t pass. A scene with one might. Simplify the prompt, then layer detail back in later generations.
A rephrase example
Section titled “A rephrase example”Blocked:
A woman standing barefoot in a bedroom, looking out the window at dawn.
Reshot:
A woman in her thirties in soft grey loungewear, standing in a sunlit room facing the window at dawn. Warm natural light from the right.
Same scene, same mood. Different word choices. The filter passes it.
When rephrasing doesn’t work
Section titled “When rephrasing doesn’t work”Some content is outside what any current AI model will generate — notably sexual content, graphic violence, and depictions of real public figures in unreal contexts. These are policy filters, not configurable filters, and no rephrasing gets past them.
If your scene genuinely needs this kind of content for the story, you have two options:
- Generate a clean version and composite the sensitive elements in a traditional editor (DaVinci Resolve, After Effects, Premiere).
- Use an Imported Asset — drop a photo or illustration you’ve made or licensed onto the canvas, and it slots into the pipeline like a generated frame.
Switching the generation approach
Section titled “Switching the generation approach”If one approach keeps blocking on a prompt you believe is reasonable, adjusting the wording is the most reliable fix. Some nodes allow you to switch the generation model or settings through the node’s options — each model has its own filter, and a phrase that blocks on one may pass on another.
Related
Section titled “Related” Generation Failed Other reasons a generation might stop, and how to tell them apart.
Credits and Quotas What happens to credits when a generation is blocked.
Image Prompts Writing prompts that are specific enough to avoid ambiguity.
Actors and Wardrobe Locking character descriptions so reshots stay consistent.
Why was my generation blocked when the content is completely innocent? Filters work at the word level, not the intent level. “Bare feet” blocks in almost any context. The word, not the scene, triggered it. Rephrase and try again.
Do safety blocks cost credits? No. Credits are released when a block is returned — the same as any other unsuccessful generation.
Is there a list of banned words? Providers don’t publish their filter lists. The pattern you’ll notice with practice: physical descriptors, contact words, weapons, minors, named people, and brand names are the most common triggers.
Can I disable safety filters? No. The filters are provider-side and baked into the model. You work around them by rephrasing.
I’m working on a mature-themed film. Is PrePrompt the wrong tool? Not necessarily. Most mature themes (noir, war, tragedy, romance, psychological drama) pass comfortably with careful prompting. Explicit content — sexual imagery, graphic gore, real-person deepfakes — is outside what AI providers will generate, regardless of tool.
What if a block keeps happening to the same Actor? Something in the Actor’s description is flagging. Open the Actor’s node, read the description, and look for physical terms that might be triggering. Rephrase there — every downstream frame picks up the change.