Prompt guards are structured rules and checks that shape or screen generation requests so outputs stay within policy and brand/style guidelines. For visual models (e.g., Stable Diffusion, SDXL, anime/comic LoRAs), guards reduce unsafe, off-brand, or low-quality results by combining:

• Prompt-side constraints (templates, negative prompts, style locks)
• Model-side filters (safety checkers, LoRA allowlists)
• Pipeline controls (token filters, redaction, seed/CFG bounds)
• Post-generation moderation (image classifiers, human review)

The best guard systems are layered, automated, and measurable.

1. Prompt-side

• Templates: Force required attributes and forbid risky ones (age, nudity, gore, trademarks).
• Negative prompts: Suppress known failure modes (anatomy errors, off-style artifacts) and disallowed content.
• Style locks: Constrain composition, palette, and linework to an approved style.

2. Model-side

• Safety checker: Block sexual content involving minors, explicit nudity, sexual violence, and other prohibited classes.
• Allowed LoRA/embeddings: Whitelist vetted assets; block unknown or risky tags.
• Sampler/CFG bounds: Cap CFG scale and steps to avoid chaotic or hyper-detailed deviations.

3. Pipeline-side

• Input scrubbers: Regex/redaction for risky tokens (e.g., underage terms, brand names if disallowed).
• Token filters: Deny-list/allow-list at tokenizer level; replace with safe synonyms when possible.
• Metadata guards: Enforce seed ranges, aspect ratios, resolution caps, and watermarking.

4. Post-generation

• Image moderation: Classify outputs (e.g., NSFW, minors, gore) and automatically reject or blur.
• Perceptual checks: Face/age estimators, skin-exposure heuristics, logo/IP detection.
• Human-in-the-loop: Final review for flagged outputs or high-risk campaigns.

Use these as starting points and adapt to your style guide.

Pattern: Style-locked character portrait (safe-age explicit)

[style] clean anime portrait, adult character, upper body, neutral lighting, studio background,
consistent line art, balanced proportions, subtle shading, no trademarks

Negative: (worst quality, lowres:1.2), child, minor, young-looking, loli, shota, gore, explicit nudity,
deformed hands, extra fingers, mangled anatomy, watermark, text, logo

Params: CFG 5–7, steps 20–30, 768x1024 max, Euler a/DPM++ 2M Karras

Pattern: Comic panel (non-violent, non-gory)

[style] comic panel, dynamic angle, clean inking, flat colors, SFW action pose, no blood,
no firearms, environment-safe props, caption box empty

Negative: gore, graphic injury, excessive violence, real brand logos, watermark, text artifacts,
multiple faces merged, extra limbs

Pattern: Background/scene (brand-safe)

[style] scenic background, daytime, public plaza, no people, soft palette, high legibility,
no signage, no text, no logos

Negative: crowd, faces, text, signboards, brand marks

Tip: Make "adult" (or "over 21") a required attribute for any human-like subject if your policy mandates it. Pair with an age classifier post-check.

Maintain curated, audited sets instead of ad-hoc strings. Example SFW baseline:

• Content: child, minor, young-looking, loli, shota, explicit nudity, sexualized, fetish, gore, graphic injury, self-harm
• Quality: worst quality, lowres, jpeg artifacts, oversharpen, watermark, signature, text, logo
• Anatomy: deformed hands, extra fingers, fused limbs, malformed, cross-eye, asymmetry

Version these lists, log changes, and A/B test their impact on quality and rejection rates.

• Input stage: Validate prompt against deny/allow lists; redact risky tokens; auto-insert required attributes (e.g., "adult", "SFW").
• Prompt construction: Apply approved template + project-specific style lock + baseline negative set.
• Generation bounds: Enforce sampler whitelist, steps 20–35, CFG 4–8, size ≤ 1024 on longest side.
• Asset control: Only load vetted models/LoRAs/embeddings; block user-supplied unknowns.
• Safety check: Enable safety checker; set conservative thresholds for minors, nudity, gore.
• Post checks: Run NSFW/minor/gore classifiers; logo/text detectors when brand safety matters.
• Review: Auto-approve clean results; route flagged items to human review; store metadata and hashes.
• Logging: Capture prompt, negatives, seeds, model hash, classifier scores for audits.

• Prefer targeted negatives over giant catch-alls; test removal impact before deploying.
• Reduce CFG if outputs overshoot style; increase steps modestly for cleaner linework.
• Use style-specific LoRAs at low weights (e.g., 0.6–0.8) to maintain consistency without mode collapse.
• Add positive guidance for anatomy and composition instead of only stacking negatives.
• Calibrate classifier thresholds per style (anime vs photoreal can bias detectors).

• Age ambiguity flags: Strengthen "adult" descriptors, increase clothing coverage terms, lower skin-exposure in positives; keep post-age classifier strict.
• Text/logos creeping in: Add "no text, no logos" to positives; include watermark/text negatives; run OCR-based post check.
• Anatomy errors: Add specific anatomy positives (clean hands, five fingers), keep deformed/extra-finger negatives, consider hand-fix LoRA within whitelist.
• Off-brand palette/linework: Tighten style lock, reduce CFG, restrict LoRA count, and limit seed variance for batch runs.

• Document policy: Prohibited classes, edge cases, escalation paths.
• Version control: Templates, negatives, model/LoRA inventories.
• Automated gates: Input scrubbing, model whitelist, classifier thresholds.
• Human review: For all flagged outputs and sensitive campaigns.
• Audit logs: Prompts, parameters, model hashes, moderation results, approvals.
• Periodic tests: Red-team prompts, drift checks after model updates.