Skip to main content
images12345
images12345
Participant
July 16, 2026
Answered

AI Hallucinations.

  • July 16, 2026
  • 1 reply
  • 20 views

If I'm paying 400 credits to create a video of 8 seconds, when I input a reference image and specify in the prompt: IMPORTANT: DO NOT CHANGE THE REFERENCE IMAGE IN ANY WAY-KEEP THE TEXT, COLOR, PROPORTIONS., I expect the AI to not change the look of the product. As it is, I spent 800 credits on video footage I can't use, because the product doesn't look exactly like the product I sell! Can you refund my credits please?

Uncle Harry’s Miracle Mouthwash 8 fl oz

 

    Correct answer Hajar Toutsi

    Hi images12345,
    Thanks so much for the detailed breakdown and the before/after examples, that really helps clarify what's happening.
     

    A couple of thoughts on how to get closer to what you're going for:

    • Negative Prompting is unsupported: Firefly's models don't currently support negative prompting (instructions like "DO NOT change"). Because models are trained to respond to direct, affirmative commands, a negation is often hard for them to interpret reliably.
      For future prompts, framing your request around what to preserve (e.g., "Keep the existing product exactly as shown...") provides much clearer guardrails for the model than telling it what not to do.
      I’ve attached a recording of a video generation of your product where all elements are kept preserved using this prompt: “Every element of the product bottle stays perfectly preserved exactly as they appear. The bottle sits centered on a clean white marble countertop, softly lit by bright natural daylight streaming from the left. The camera slowly pushes in toward the bottle in a smooth, cinematic dolly movement, keeping the label facing the camera at all times. Delicate water droplets glisten on the bottle's surface, catching the light. Fresh mint leaves rest gently around the base of the bottle. The background stays softly blurred with a clean, minimal spa-like aesthetic in cool white and pale turquoise tones. The mood feels fresh, clean, and premium.”

    • Text Generation & Product Angles: Accurately generating specific text is a known, general limitation of current AI. Because your product is highly text-heavy, providing just a single reference image leaves "blind spots." When the AI doesn't have all the necessary angles of the physical product, it creatively predicts the missing information to fill in the gaps which results in the text and design deviations you are seeing.
      Providing reference images from multiple angles is the best way to maintain true text and design likeness.


    Unfortunately, credits are the cost of each generation, so they can’t be refunded for unwanted outcomes.
    That said, we’d love to help you get closer to the results you want, just share the details of what you have in mind and we’ll be here to guide you.

    Hajar
    [Moved to questions]

    1 reply

    Hajar Toutsi
    Community Manager
    Hajar ToutsiCommunity ManagerCorrect answer
    Community Manager
    July 16, 2026

    Hi images12345,
    Thanks so much for the detailed breakdown and the before/after examples, that really helps clarify what's happening.
     

    A couple of thoughts on how to get closer to what you're going for:

    • Negative Prompting is unsupported: Firefly's models don't currently support negative prompting (instructions like "DO NOT change"). Because models are trained to respond to direct, affirmative commands, a negation is often hard for them to interpret reliably.
      For future prompts, framing your request around what to preserve (e.g., "Keep the existing product exactly as shown...") provides much clearer guardrails for the model than telling it what not to do.
      I’ve attached a recording of a video generation of your product where all elements are kept preserved using this prompt: “Every element of the product bottle stays perfectly preserved exactly as they appear. The bottle sits centered on a clean white marble countertop, softly lit by bright natural daylight streaming from the left. The camera slowly pushes in toward the bottle in a smooth, cinematic dolly movement, keeping the label facing the camera at all times. Delicate water droplets glisten on the bottle's surface, catching the light. Fresh mint leaves rest gently around the base of the bottle. The background stays softly blurred with a clean, minimal spa-like aesthetic in cool white and pale turquoise tones. The mood feels fresh, clean, and premium.”

    • Text Generation & Product Angles: Accurately generating specific text is a known, general limitation of current AI. Because your product is highly text-heavy, providing just a single reference image leaves "blind spots." When the AI doesn't have all the necessary angles of the physical product, it creatively predicts the missing information to fill in the gaps which results in the text and design deviations you are seeing.
      Providing reference images from multiple angles is the best way to maintain true text and design likeness.


    Unfortunately, credits are the cost of each generation, so they can’t be refunded for unwanted outcomes.
    That said, we’d love to help you get closer to the results you want, just share the details of what you have in mind and we’ll be here to guide you.

    Hajar
    [Moved to questions]

    Terms & ConditionsAccessibility statement
    Using the community Terms of use Privacy Cookie preferences Do not sell or share my personal information ad choicesAdChoices