Image-to-video vs text-to-video

Same price, different risk

On most platforms — including Sora 2, where image conditioning is explicitly free — both input modes cost the same per second. What changes is the failure rate: anchoring generation to a real image removes the model's biggest source of error (inventing your subject wrong) and typically cuts retake ratios 30–50% on anything specific.

When each wins

• Image-to-video: products, real people, brands, locations, listing photos, any 'this exact thing' shot. The accuracy isn't just cheaper — it's often the legal/commercial requirement (product demos, real estate).
• Text-to-video: concepts, atmospheres, things that don't exist yet, and exploration — where the model's imagination is the point and a reference image would only constrain it.

The hybrid that pros use

Generate a still first (image models are cheaper and faster to iterate), perfect the frame, then animate the winner. You get text-to-video's creative range with image-to-video's failure rate — and frame iteration costs cents instead of dollars. For pipelines, this two-step is the highest-leverage habit after draft-tier discipline.