Critical Bug: Video Generation Model Bypasses Pre-Flight Directives, Ignores Reference Images, and Hallucinates Unauthorized Objects

Hello Support Team and Community,

We are facing a critical, persistent system-level failure in the video generation engine regarding prompt adherence, custom image constraints, and verification protocols.

Core Issues Encountered:

Total Bypass of System Directives & Pre-Flight Checks: System-level instructions, automated pre-flight verification routines, and saved configuration parameters are completely ignored by the processing engine.

Ignored Image Reference Constraints: When providing exact reference frames/images, the video model fails to preserve the source pixels accurately, instead overriding them with a rigid default template.

Unwanted Object Hallucination: Despite strict negative constraints and clear textual boundaries restricting visual elements exclusively to a defined 5-point aesthetic framework (Bindi, Kajal, Pink Lipstick, Sindoor, and Mangalsutra), the engine repeatedly injects unauthorized accessories (such as nose rings) into the output frames.

Resource Exhaustion: Because the model continuously overrides instructions and generates non-compliant assets, generations repeatedly result in errors, burning through daily rate limits without yielding usable production output.

Expected Behavior: The video generation pipeline must strictly and immutably respect user-provided reference images, active prompt parameters, and verification flags without defaulting to unauthorized templates or hallucinating unrequested attributes.

Please escalate this prompt-adherence and template-override bug to the engineering team for a backend fix.