How to Choose an AI Video Creation Route: Prompt-Driven or Reference-Image-Driven? NadouPro’s Logic and Practice

Anyone making AI video will eventually run into this fork: for this creation, should you write the idea as a prompt and let the model develop it, or upload reference assets and let the model carry them forward? The two routes have completely different capability boundaries, suitable scenarios, and writing methods. Choose the wrong route, and rework costs rise many times over. This article breaks down the two routes clearly and gives specific selection methods in NadouPro.

The essential difference between the two routes

Prompt-driven: Write the scene direction in text, and the model is responsible for translating it into images. It offers the highest creative freedom and is suitable for projects where ideas are still being explored and visual assets have not yet been finalized.

Reference-image-driven: Use existing images, character settings, and product photos to guide generation, and the model is responsible for making the new image continue the existing visual information. It offers the highest certainty and is suitable for projects where visual assets already exist and consistency must be maintained.

There is only one criterion: Is the reference material part of the creative requirement, or merely optional inspiration? If a client’s brand character, a product’s real packaging, or the protagonist of serialized content must be accurately reproduced, use the reference-image-driven route; if everything is imagined from scratch, use the prompt-driven route.

Prompt-driven: how to write the scene clearly

In prompt-driven projects, the text has to carry all creative instructions. A qualified scene description includes four things:

Who is present: Clearly write each character’s appearance, clothing, and position. Multi-character scenes especially need clear labels for everyone—what Character A does and what Character B does, with actions and lines assigned separately, to avoid the model mixing up behaviors.

What happens: Actions should be specific enough to execute. “Two people argue” is less effective than “one person steps forward, the other takes half a step back, and spreads both hands.”

How the camera moves: Shot size, camera position, and movement direction. When the camera language is vague, the model will make its own choices, and the result becomes unpredictable.

Emotion and sound: Overall mood, dialogue content, and ambient sound. All video models in NadouPro support audio-video synchronization, so sound can be written into the same generation instruction; for scenes that require high dialogue precision, create the lines separately with dialogue dubbing and then composite them.

The prompt-driven route places high demands on the model’s semantic understanding. In NadouPro, Seedance 2.0 focuses on extremely strong semantic understanding, while Seedance 2.5 has strong instruction understanding and stable camera scheduling, making them the preferred choices for this route.

Reference-image-driven: how to provide assets properly

In reference-image-driven projects, the decisive factor is the organizational quality of the reference assets. NadouPro video nodes provide multiple ways to carry them forward:

Multi-reference mode: Input multiple reference images at the same time, with characters, scenes, and props each in their own place. The key is clear division of labor—which image defines appearance, which defines clothing, and which defines the environment. Ambiguous inputs will interfere with each other.

First and last frames: When both the starting image and ending image are given, the model fills in the process between them. This is suitable for transition design and action continuity.

Motion control: Use reference motion to drive character movement, suitable for scenes such as dance and martial arts that have clear action templates.

Typical reference-image-driven projects include brand advertising (where the product and endorsement image must be recognizable), serialized narratives (the same character appears across episodes), and series content (fixed visual assets are reused repeatedly). The standard workflow for this type of project in NadouPro is: first use multi-panel grids on the image side to produce character look-development assets (full-body three-view character sheet, three-view character face sheet, nine-grid character expressions), use the finalized look images as references throughout every video generation, and manage all shots on one canvas.

In practice, the two routes work together

Mature projects rarely use only a single route. A typical combination is:

Use prompt-driven creation for early exploration: When the direction is undecided, use text-to-video to quickly test multiple visual versions at the lowest cost.

Switch to reference-image-driven after assets are finalized: After characters and scenes are selected, produce finalized look assets, and use multi-reference for all subsequent generations.

Use editing nodes in the finished-film stage: Multi-shot materials enter editing nodes for stitching and online preview. Segments that require precise camera-position scheduling are produced separately with the 3D director’s table.

Take a typical scenario as an example: for a branded short drama project, the first week uses prompt-driven creation to explore the protagonist’s image and the scene tone; after finalization, character three-view sheets and a four-grid scene mood board are produced; the following twenty shots are all generated using the finalized look assets as references, making character consistency stable and controllable; finally, the film is output through the editing node.

How to choose a model: by route plus scenario

All twenty-two video models in NadouPro support audio-video synchronization, matched by route and scenario:

Preferred for prompt-driven creation: Seedance 2.0 (extremely strong semantic understanding), Seedance 2.5 (strong instruction understanding and stable camera scheduling).

Preferred for reference-image-driven creation: Seedance 2.5 (all-around reference support, strong in both quantity and quality), Kling 3.0 Omni (multimodal input).

Multi-shot narrative: Wan 3.0 (intelligent multi-shot), MiniMax H3 (multiple shots output together), Happy Horse 1.1 (intelligent multi-shot), Vidu Q3 (supports video storyboards).

Special needs: For anime style, choose MiniMax-2.3 (best anime expressiveness); for high physical plausibility requirements, choose Bach 1.0.

FAQ

Q: Can the same project switch routes midway?
Yes, and it is common. Use prompt-driven creation during exploration, then switch to reference-image-driven creation after assets are finalized. The timing for switching is “the visual direction has been set and production is entering scale,” and the switching action is to organize the selected images into a reference asset library.

Q: In reference-image-driven creation, will the model copy the reference image exactly?
No. Reference images guide identity, style, and key details, while composition, action, and lighting are still determined by the prompt. To maximize fidelity, state clearly in the prompt: “keep the clothing, hairstyle, and scene elements from the reference image unchanged.”

Q: How should multi-character dialogue scenes be written so they do not become chaotic?
Three rules: write each character’s appearance description fully and use fixed wording; place actions and lines immediately after the corresponding character description; create dialogue separately with dialogue dubbing and then composite it, which is more controllable than relying on audio-video synchronization during generation.

Q: How should audio-video synchronized sound and dialogue dubbing work together?
Audio-video synchronization handles atmosphere layers such as ambient sound, sound effects, and background music; dialogue is created separately with dialogue dubbing, with stable voice color and controllable emotion, then composited in the editing node. Iterate the two layers separately, so changing dialogue does not require regenerating the visuals.

Q: How can I tell whether a model fully understands my prompt?
Use a small-sample test: generate once with the same prompt on each candidate model, and compare four points—whether the characters appear as described, whether the actions are executed, whether the camera follows the instruction, and whether any details are missing. For semantically complex scripts, test Seedance 2.0 first.

For more features, you are welcome to try NadouPro: NadouPro

Contains AI-generated information that may be incorrect. Please verify independently.