Most of the misleading output here is not a model problem, it is a pipeline problem: tools generate visuals from a product title alone, so the render invents fabric, size, and packaging the listing never claimed. The practical fix is to constrain generation to seller-supplied assets for anything that shows the product, and keep fully synthetic shots for b-roll and framing. A per-scene review gate before publish catches most of it, which each scene part editable shot by shot in Maqtai rather than forcing a full regenerate. Platform side, Shopee and TikTok could require a provenance flag on ad creatives and treat generated voiceover claims as seller claims for takedown. It also allows to edit just one scene from video if you don't like one scene instead of regenerating the whole video.