Character consistency

Generated clips drift. Ask the same model for the same person twice and you get cousins. This is the method that actually holds identity across a sequence, and it costs one extra image generation, not a custom model.

The problem, precisely

Each clip is generated independently. A text prompt — however detailed — does not pin a face. "A woman in her late twenties with shoulder-length brown hair" describes thousands of people and you will meet several of them.

The method: chain the frames

clip 1  ─ generate ─► extract last frame ─► upload_image
                                                 │
clip 2  ◄─ generate from that frame as the first frame
                                                 │
clip 3  ◄─ …and so on

Identity holds perfectly, because clip 2 does not re-invent the person — it starts from a picture of them.

1. Make clip 1

Any text-to-video model. This one establishes the person, the wardrobe and the room, so spend your prompt effort here.

generate_video { "model": "seedance-2.0-t2v", "prompt": "…", "duration": 5 }

2. Take a frame out of it

Download the MP4 and pull a frame with ffmpeg — late in the clip, where the pose has settled:

ffmpeg -ss 4.5 -i clip1.mp4 -frames:v 1 frame1.png

3. Put it back in as the first frame

upload_image { "image_base64": "…" }

generate_video {
  "model": "seedance-2.0-frames",
  "image_url": "<the uploaded frame>",
  "prompt": "<what happens next>",
  "duration": 5
}

The *-frames models take the image as the literal first frame. seedance-2.0-frames is the quality option; seedance-2.0-fast-frames is cheaper and silent, which is fine for B-roll where nobody speaks.

The trap: the first frame anchors the expression

This is the part that surprises people, and it cost us a render to learn.

If clip 1 ends on a tearful face and clip 2's prompt asks for a suspicious squint, you get a tearful squint. The first frame does not just pin the identity — it pins the emotion and the pose, and the prompt cannot override it.

To change expression or posture between clips:

frame from clip 1 ──► generate_image (new expression, frame as reference)
                              │
                              └──► generate_video  *-frames  from THAT

You are generating a new keyframe of the same person wearing the new expression, then animating it. One extra image call, and the sequence works.

The same trick holds a product label

Product text mutates between shots unless it is in a reference image. Generate each shot's keyframe with generate_image using one product photo as the reference, then animate the keyframes. We have run four shots this way with the label correct in all four.

Rules that made the difference:

• One product photo as reference, not several. Two photos give the model

objects to juggle — bottles vanish, hands swap.

• Say the product stays in the same hand for the whole clip.
• One simple motion per clip — tilt to camera, pat onto cheek. Not a

sequence.

When not to bother

If you want a finished ad rather than a sequence you assemble yourself, use create_ad_video — it handles the creator and product internally and returns one rendered MP4. This skill is for multi-clip work where you are cutting the clips together in your own editor.

What fails

• Re-describing the person in text for each clip. This is the thing that

does not work, and it is everyone's first attempt.

• Extracting a frame mid-blink or mid-word. Scrub for a clean one.
• Expecting the prompt to beat the first frame. It does not. Make a new

keyframe.

• Several reference images at once. Fewer references, more consistency.

Spending rules

• Every generate_video and generate_image call spends credits. quote

first — it charges nothing and returns exactly what the call will cost.

• A 4-clip sequence is 4 video calls plus the keyframes. Price the whole

sequence before the first one.

• Failed runs refund in full.