How I Make Realistic AI Videos with Seedance 2.5 - Full Guide

Summary

Dan Kieft demonstrates a workflow for creating realistic AI images and videos using Higgsfield, which sponsored the video. His main advice is to test multiple models and build realism through deliberate prompting and camera behavior rather than relying on a single tool or prompt.

Model selection

Image models: Kieft recommends testing the same prompt with several models, since each has different strengths. He highlights Flux 2 Pro, GPT Image 2.5 Flare, Nano Banana models, SeaDream 5.0, and Soul as options worth trying. Some models may be better suited to realism, particular visual styles, or following text instructions.

Video models: In his comparisons, Kieft considers Seedance 2.5 at 1080p the strongest for realistic footage, though relatively expensive. Flux 3 is a cheaper alternative that can approach much of the result. He suggests using Seedance’s 480p draft mode to test multiple generations, then upscaling a preferred result to 1080p. In his test, Minimax was less realistic; he also mentions Kling as another model to try.

These are the creator’s impressions based on the examples shown, not universal rankings.

Prompting guide: The CTLT method

For image prompts, Kieft teaches CTLT:

1. Camera: Describe the lens, framing, angle, and photographic format.
2. Tone: Set the mood, atmosphere, and emotional quality.
3. Lighting: Specify the light source and how it falls on the subject.
4. Texture: Add physical details and describe how materials interact with the environment.

He recommends using an LLM such as ChatGPT or Claude to turn a basic idea into a structured prompt. Reference images can also help guide the result.

Style and creative direction

Small changes in style instructions can create substantially different images, such as a disposable-camera flash, camcorder footage, phone snapshot, or documentary look. Kieft suggests intentionally using imperfections—such as grain, blur, compression, or low resolution—when they support the desired realistic aesthetic.

He also describes reusable prompt structures and suggests asking an LLM to create a Markdown “skill” file from a successful prompt and image so the style can be reused.

Developing stronger ideas

Kieft uses contrast to explain why an image may catch attention:

• Level 0: The subject is in an expected setting.
• Level 1: The subject is placed in an unexpected situation.
• Level 2: Add unexpected styling or personality.
• Further contrast: Add an emotional element that gives the image a feeling or story.

Video realism techniques

To make AI footage seem human-filmed, Kieft recommends prompting for camera behavior such as:

• Handheld movement and natural shake
• Crash zooms and whip pans that react to events
• Voyeuristic or candid framing
• Walking backward with uneven movement
• A person or object passing in front of and blocking the frame
• Exposure adjustments when moving between lighting conditions
• Focus hunting or difficulty refocusing

He suggests combining several techniques and thinking like a director when describing the shot.

Guides and resources

Kieft says he provides prompts, example assets, a CTLT template, a guide to seven ways of capturing an idea, and reusable Markdown files through his linked community and resources. He also points viewers to a separate video about saving credits on AI video generation.

Speaker and sources

• Main speaker: Dan Kieft, presenting the tutorial and his own model comparisons
• Platform: Higgsfield, identified as the video sponsor
• Other tools referenced: ChatGPT and Claude for prompt development, along with the image and video-generation models discussed above.