Disable Safety Checker for FAL Image Generation

[daojonesceo]

Is your feature request related to a problem? Please describe.

Many image generation services such as FAL, Stable Diffusion, etc. have "safety checkers" that ensure prompts are appropriate.
These have several problems:

1. Additional latency: These checks are done via an LLM, which can add unnecessary latency.
2. Unpredictable / incorrect outputs: Since the checks use LLMs, their output can be overly restrictive or random, causing unpredictable behavior.
3. Not necessary: Some use cases don't require the safety checker.

In particular, we at daojones.ai (@DAOJonesPumpAI on X.com) have found this is overly restrictive and have written a custom image generation application that disables it.

Describe the solution you'd like

We want to set the FAL configuration for our image requests.
We need to set enable_safety_checker to false while also setting safety_tolerance to 6 (the most tolerant setting).
Since these options are not uniformly supported by the different APIs, we will just add an "escape hatch" configuration imageSettings.fal.

Describe alternatives you've considered

1. Turning safety checker off globally: This would work for our use case but could be a breaking change for other use cases.
2. Adding a global allowNsfw flag: This requires reconciling this with other APIs and determining what settings it would correspond to.
3. Adding the enableSafetyChecker / safetyTolerance flags at the top level of imageSettings: this would pollute the global namespace.

Additional context

[github-actions]

This issue has been automatically marked as stale due to inactivity.