FAQ (Three-in-One): Jump Cut · Twin Characters · Quality Degradation

---

1. Jump Cut at Seam

Typical Issue

When using the video extension feature to generate a new video and then stitching it with the original, there may be visual jumping or regression at the seam.

Solution

Currently, the recommended approach is post-editing repair. Fundamental improvements will be addressed in future model iterations.

Comparison

!Jump cut issue at seam

Before optimization: sudden frame jumping or content regression at the seam (5s, 20s)

Comparison

• Before optimization:
• After optimization:

Editing Steps (Using CapCut or Similar Software)

• Import the videos to be stitched into CapCut
• At the first seam, trim 6 frames from the end of the preceding video
• At the same time, trim 1 frame from the start of the following video
• Repeat the above steps for all stitching points
• Export and check the smoothness of the stitched video

Advanced Tips

Even after frame alignment, subtle jumping may still remain. When generating continuation videos, it is recommended to end at a transition or cut moment, and start the next segment from the new scene after the cut. This can further reduce the sense of jumping.

---

2. Twin Characters Issue

Typical Issue

In scenes with multiple characters where a three-view reference image of a character is provided as input, the generated video may show two identical characters in the same frame.

Root Cause Analysis

• The description of the character subject in the prompt is unclear, making it hard for the model to distinguish between different characters
• Using three-view / multi-view character images as reference material can confuse the model's character recognition

Solution

Currently, it is not possible to 100% avoid the twin characters issue. The following methods can reduce the likelihood of occurrence:

#### 1. Clearly Define Subject-Character Association

Clearly define each character's role in the prompt and specify the mapping between the character and the reference image.

Example:

Zhang San (corresponding to Image 1) throws the green bankbook towards Li Si (corresponding to Image 2), who is standing.

#### 2. Add Global Constraint Instructions

Append a fixed constraint at the end of the prompt:

Throughout the video, do not allow characters with identical appearance, clothing, or accessories.
Do not generate clone-like duplicates or twin effects.
In each frame, keep only the corresponding single character; do not produce repeated copies of characters.

#### 3. Optimize Reference Material

Use single-person photos as character reference images whenever possible; three-view or multi-view materials are not recommended.

#### Comparison

• Before optimization:
• After optimization:

#### 4. Streamline and Optimize the Prompt

Do not use entire scripts as prompts. Excessively long and redundant copy can confuse the model's understanding.

---

3. Quality Degradation

Typical Issue

When using a model-generated video as input material for continuation/extension, quality degradation occurs. Multiple rounds of continuation compound this degradation, especially in facial areas where blotchy color artifacts may appear.

Solution

The following methods can help mitigate quality degradation for now. Fundamental improvements will be addressed in future model iterations:

#### Method 1: White-Model Video Approach

Convert the original video into a white-model video using Seedance 2.0, then use it as input for continuation.

Reference prompt:

Convert the video into a white 3D model. All characters are rendered as pure white 3D models,
no colors, no textures, no shadows, pure white background, stable structure, smooth motion.

Comparison:

• Before optimization:
• White-model fix:

#### Method 2: High-Quality Reference Images

Prioritize high-resolution images as reference material.

#### Method 3: Control Continuation Count

Keep the number of video continuation rounds reasonable to avoid compounding degradation.