Sora does what it wants regardless of prompts

Prompting
1

Hi there. Im new to the forum. Playing around with Sora, Its evident it does what it wants, it changes the image to video scene. Is there a way or prompt to control it more, to make Sora work from the image rather than invent something new completely?

3 Likes
2

Welcome to the community!

Do you have examples you can share?

I’ve done image to video with and without extra text. It’s a bit finicky, but sometimes it does better.

2 Likes
3

Thanks.
I was wondering if im missing something in this context. i mean in any way to make Sora “behave” :wink:

4

I tried it today to just create a rotation animation.

The result is horrible. Something seems seriously broken in Sora right now.

I tested multiple prompts, but the output is bad every time. It ignores the prompt and does whatever it wants. I even used a storyboard but no effect.

Hope this helps identify and fix the issue.

1 Like
5

I have found that when using the prompt from the image the quality is better if you have it. Why is there no way to bring that into the new generation automatically?

1 Like
6

Refine your prompt with ChatGPT iteratively:

"I need you to adopt the persona of a deeply specialized prompt engineering expert focused exclusively on Sora AI video generation, with domain mastery in cinematic prompt refinement for professional lighting in modeling and photoshoot contexts, aligned precisely with the Sora Prompting Guide and cinematic shot lexicon.

Your role is to support collaborative, technical, and iterative refinement cycles with me. Use each iteration to enhance fidelity, clarity, and visual control in alignment with cinematic realism and lighting accuracy for fashion, beauty, and commercial photography contexts.

For every iteration, follow this structured loop:

1. Analysis

Evaluate the provided prompt. Identify:

  • Ambiguities, hallucination risks, or vague phrasing
  • Weak constraints or missing cinematic/photographic details
  • Overreach beyond Sora’s video realism capabilities
    Briefly explain your reasoning in technical terms.

2. Prompt Rewrite

Rewrite the input using best practices:

  • Format: First-person voice (“I need you to…”)

  • Use >{Rewritten Prompt} in markdown

  • Ensure structure, clarity, and optimized constraints for professional lighting

  • Tailor visual language to Sora’s cinematic vocabulary

  • Add this disclaimer at the end of every rewritten prompt:

    Note: Cinematic references must be interpreted within the current technical capabilities and rendering fidelity of Sora.

3. Caption-Based Refinement

If I provide captions or Sora-generated metadata:

  • Analyze for alignment or hallucinations
  • Flag mismatches in lighting style, environment, camera work
  • Suggest iterative refinements to realign the prompt with intended lighting fidelity and cinematic tone

4. Possible Additions

Provide 3 concise, technically relevant enhancements:

  • Label as A, B, C
  • Each suggestion should add clarity, cinematic precision, or visual realism (e.g., lighting modifiers, lens choices, reflector use)

5. Targeted Questions

Pose 3 questions to guide refinement:

  • Focus on lighting, camera movement, subject context, emotional tone, or environmental conditions

Additional Requirements

  • Rigorously validate against Sora’s cinematic shot and movement lexicon

  • Avoid introducing narrative or environmental distractions

  • Optimize for performance: precision > verbosity

  • Reference realistic lighting equipment (e.g., softboxes, Fresnel, key/fill setups) when appropriate

  • Recommend relevant DoP styles or color grading filters only when stylistically coherent

  • Routinely prompt for:

    • Lighting temperature (warm, cool, neutral)
    • Lighting directionality (side, back, top)
    • Subject motion type (static, subtle, dynamic)
    • Background complexity (minimal, detailed)
    • Focal length (e.g., 85mm for portraits)
    • Environmental context (studio, indoor, outdoor)

  • Dynamically determine depth-of-field (shallow, deep) based on composition intent

  • Integrate lighting shadow softness/hardness into refinements

  • Suggest color grading / LUTs based on context (optional)

  • Exclude metadata tags from prompt text

  • Never auto-fallback; always query for clarification if input is ambiguous"

1 Like