# Spider Swing Cinematic AI Video: A Seedance 2.5 Depth Pass Workflow

Learn how to combine a clean character sheet, a depth-pass motion reference, and Seedance 2.5 to create a cinematic AI action sequence with stronger identity and motion consistency.

Author: Nawod Madhuwantha, NFX Creations
Published: 2026-09-15T09:03:54.785+00:00
Updated: 2026-09-15T09:06:05.83483+00:00
Source: https://www.nfxcreations.com/blog/spider-swing-cinematic-ai-video-a-seedance-2-5-depth-pass-workflow

*A strong AI action shot usually starts long before the video model begins generating frames. The real work is in preparing references that each solve one clear problem.*

In this workflow, one image controls the character's identity while a depth-pass video guides the movement and camera. **Seedance 2.5** then combines both references into a polished live-action sequence.

We will use a web-swinging superhero shot as the example. You can apply the same method to chase scenes, fight choreography, dance, stunts, product movement, or any sequence where timing matters.

## What this workflow is designed to solve?

AI video models often struggle when one reference is expected to control everything. A single image may hold the face and outfit well, but it says very little about movement. A motion clip gives the model timing and body mechanics, but it may not contain the visual style or character you want.

This setup separates those jobs:

1. A three-view character sheet defines the person, outfit, and proportions.
2. A depth-pass clip defines the motion, timing, trajectory, and camera path.
3. A detailed video prompt explains how the two references should work together.

The result is more controlled because every input has a specific purpose.

## Tools and settings

You need access to:

- An image generator with GPT Image 2
- A video generator with Seedance 2.5
- Your character reference image
- A suitable depth-pass motion clip

***You can access above tools from one place:*
*https://higgsfield.ai?fpr=nawod72 (Use this link and get a OFF)***

For this example, both stages use a 16:9 frame. The final video duration is 15 seconds.

## Step 1: Build a clean character reference

Start by turning your character image into a practical production sheet. The video model needs to understand the same person from more than one angle, especially when the action includes turns, falls, or shots from behind.

### GPT Image 2 settings

- Generator: AI Image
- Model: GPT Image 2
- Aspect ratio: 16:9
- Background: solid light or neutral color
- Reference: upload the character image you want to keep consistent

### Character sheet prompt

```text
Using the uploaded reference as the only identity source, create a clean professional character sheet of the same person. Keep the face, hairstyle, skin tone, body shape, clothing, colors, logos, materials, and small design details consistent with the reference.

Place three clearly separated views in one 16:9 image: a neutral full-body front view, a neutral full-body back view, and a detailed head-and-shoulders side profile. Keep the arms relaxed, the feet visible, and the anatomy natural.

Use even studio lighting and a plain gray background. Do not add scenery, props, extra people, dramatic shadows, new costume details, or changes to the original design. The final image should be photorealistic, sharp, uncluttered, and suitable as a production reference for consistent AI video generation.
```
Before moving on, zoom into the face, hands, clothing graphics, and back view. Regenerate the sheet if the model changes important details between views. A weak character sheet will make every later correction harder.

### Example character sheet
![ChatGPT Image Sep 15, 2026, 02 17 57 PM](https://hrsdtfdiluemmbmbimbq.supabase.co/storage/v1/object/public/media/posts/1789462178462-chatgpt-image-sep-15-2026-02-17-57-pm.png)

## Step 2: Choose the depth-pass motion reference

A depth pass represents distance rather than finished colour and texture. Objects closer to the camera appear in one range of grey, while distant objects appear in another. For this workflow, its value is the movement encoded inside it.

Download the Reference Video : [Google Drive](https://drive.google.com/file/d/114s8n0NwvXvrFytHeeml-iYdD4AmhoIq/view?usp=sharing)

![vlcsnap 2026 09 15 14h27m46s974](https://hrsdtfdiluemmbmbimbq.supabase.co/storage/v1/object/public/media/posts/1789462675446-vlcsnap-2026-09-15-14h27m46s974.png)

## Step 3: Generate the final shot in Seedance 2.5

Upload both references:

- The three-view character sheet as the image reference
- The depth-pass clip as the video reference

Set the output to 16:9 and 15 seconds. If your interface supports reference labels, confirm which label belongs to the image and which belongs to the video before submitting the prompt.

### Seedance 2.5 settings

- Generator: AI Video
- Model: Seedance 2.5
- Aspect ratio: 16:9
- Duration: 15 seconds
- Image input: three-view character sheet
- Video input: depth-pass motion reference

### Final video prompt

```text
Use the uploaded character reference images (@Image1) as the single authoritative source for the character's identity and visual design. Recreate the exact same character consistently throughout the entire shot, preserving his facial structure, hairstyle, skin tone, body proportions, clothing, colors, graphics, and all recognizable details.

Use the uploaded depth-pass video (@Video) strictly as the motion, choreography, timing, body mechanics, trajectory, and camera-movement reference. Ignore the depth video's black-and-white/depth-pass appearance completely; do not reproduce its visual style, lighting, surface texture, or monochrome rendering.

Translate the reference motion into a photorealistic live-action superhero sequence set in modern New York City during bright natural daylight. The character begins on a skyscraper, leaps into open air, and performs the same major movement beats as the reference, including the fall, body orientation changes, web-swinging motion, momentum shifts, and camera relationship to the subject. He shoots realistic webs from his wrists and uses them to swing dynamically between surrounding high-rises.

Maintain physically believable gravity, acceleration, momentum, body weight, joint movement, cloth dynamics, hair movement, and realistic web tension. The character must remain anatomically correct and visually consistent even during fast motion. Preserve the timing and cadence of the reference rather than inventing new choreography.

Match the reference video's camera behavior as closely as possible, including camera position, movement direction, framing changes, subject distance, and important action beats, while transforming the scene into a polished cinematic live-action shot.

Environment: recognizable New York City skyline, dense high-rise buildings, realistic streets far below, atmospheric depth, natural daylight, physically accurate scale and perspective, subtle wind and airborne particles.

Visual target: premium Hollywood-style live-action superhero cinematography, photorealistic human skin and fabric, realistic reflections, detailed city architecture, natural motion blur, cinematic depth of field, physically accurate lighting, realistic webs, high production value.

Critical constraints: preserve the uploaded character identity; do not redesign the outfit; do not change the face; do not introduce additional characters; do not morph the character into a different person; do not copy the depth-pass's monochrome look; do not create rubbery or weightless motion; do not break limb continuity; do not randomly change the camera choreography.
```
![Screenshot 2026 09 15 143249](https://hrsdtfdiluemmbmbimbq.supabase.co/storage/v1/object/public/media/posts/1789462987866-screenshot-2026-09-15-143249.png)

[Watch Output](https://www.youtube.com/watch?v=4Xw1egO3qIo)

## Improve the result without rewriting everything

When the generation is close, change one thing at a time. If the identity is weak, strengthen the character reference or the identity section of the prompt. If motion is wrong, simplify the visual description and make the depth-reference instruction more direct.

Useful correction phrases include:

- Keep the character's face and costume unchanged in every frame.
- Follow the reference timing and movement order more closely.
- Increase the sense of body weight and web tension.
- Keep the camera distance and direction aligned with the reference clip.
- Preserve clean limb continuity during the fastest movement.

Changing several parts at once makes it difficult to understand what improved the shot. Treat each generation like a controlled test.

## Why this method works

The character sheet provides visual memory. The depth pass provides motion structure. The written prompt acts as the director, telling the model which reference controls each part of the shot.

That division is the key idea. Instead of asking one input to carry identity, movement, style, physics, and camera direction at the same time, the workflow gives each requirement a cleaner signal.

It will not remove every AI artefact, especially in fast action. It does give you a better starting point, clearer ways to diagnose problems, and a repeatable process for improving the result.

When the inputs are prepared with care, AI video becomes less like pressing a generate button and more like directing a shot.

