How to use Seedance 2.5: frames, characters, and camera control

Control the start and end of a shot with first & last frame

First & last frame lets you upload two images: one that becomes the first frame of your video, and one that becomes the last. Seedance 2.5 fills in everything between them. Instead of hoping the model drifts toward the ending you imagined, you hand it the destination and it plots the route.

This is the single most useful feature in 2.5 for most creators, because it converts "prompt and pray" into something closer to directing.

When to use it

Use first & last frame any time the movement itself is the point of the shot:

• Transitions. A close-up of snow falling on a glove as the first frame, a wide shot of a neon-lit city street at night as the last frame. The model has to figure out the camera pull-back and scene transformation — and it's surprisingly good at it.
• Product reveals. Frame one: the product hidden under a cloth or in shadow. Last frame: fully lit, front and center. Great for ad-style clips.
• Before/after. A messy desk → a clean one. A seedling → a blooming flower. Day → night over the same skyline.
• Character entrances and exits. Character off-screen (or facing away) in the first frame, mid-scene and facing camera in the last.

If you only have one strong image, use it as the first frame and describe the motion in text. Two frames is when this feature really earns its keep, though.

Step by step

1. Prepare your two images first. They don't need to be AI-generated — photos, stills from previous clips, or Midjourney/Seedream images all work. Aim for the same aspect ratio as your output video.
2. Open the video generation interface in Jimeng, Doubao, or Dreamina, and select the Seedance 2.5 model.
3. Upload both images in the reference/material panel and assign one as the first frame and the other as the last frame. Keep any text prompt you write focused on the motion between the two frames, not on describing the images themselves — the model can see them.
4. Write a motion prompt. Be specific about what changes and how the camera behaves. Compare:
   CopyVague: "a winter scene becomes a city" Better: "slow camera pull-back from the snow-covered glove, snowfall dissolves into neon rain, reveal a wide city street at night"
5. Generate and review. Watch the middle of the clip, not just the endpoints. The first and last frames will be right by construction — the question is whether the journey between them looks intentional.
6. Iterate on the prompt before changing the images. Most bad results come from an underspecified transition, not from the frames.

Tips and common mistakes

• Keep visual continuity between the two frames. Two frames with similar lighting, color palette, or a shared element (the same character, the same red umbrella) give the model a thread to follow. Two completely unrelated images force it into a hard cut or a morph — sometimes you want that, often you don't.
• Don't describe the frames in the prompt. Wasted tokens. Describe speed, camera direction, and what changes ("the fog clears," "she turns around," "the camera orbits left").
• Match the mood of the motion to the duration. If you're generating a short clip, ask for one clear action. "Camera pulls back AND seasons change AND a character walks through" is three shots' worth of motion crammed into one.
• Weird morphs usually mean the frames are too far apart. If the model melts one scene into the other in an ugly way, add an intermediate cue in the prompt ("cut through falling snow," "pass behind the pillar") to give it a transition device.
• Use your last frame as the next clip's first frame. This is the backbone of longer sequences — more on that in the chaining section below.

Keep the same character across shots with character consistency

Character consistency lets you upload reference images of a character so the same face, outfit, and proportions carry across every shot you generate. Seedance 2.5 accepts up to 50 reference assets per pass (30 images, 10 video clips, 10 audio clips), and you can dedicate a chunk of those image slots to your character's face, outfit, and angles.

If you've ever made an AI video where the protagonist's face quietly changes between shot 2 and shot 3, you know why this matters. Consistency is the difference between "a collection of clips" and "a video."

When to use it

• Multi-shot storytelling. Any narrative with the same person in more than one shot.
• Series content. A recurring mascot, virtual influencer, or spokesperson across many videos, not just one.
• Product demos with a presenter. Same host, different scenes.
• Combining with first & last frame. The two features stack: reference images keep the character stable, while your frame images control the staging.

Step by step

1. Build a small reference set for your character. 3 to 6 good images usually beat 30 mediocre ones. Include: a clear front-facing shot of the face, a 3/4 angle, a full-body shot, and ideally one image in the outfit you want for this video.
2. Keep references clean. Plain or simple backgrounds, consistent lighting, and no other people in frame. Every distraction in a reference image is something the model might "learn."
3. Create an "Element" from your references. In Dreamina, open Elements in the left sidebar (主体 in Jimeng), create a new Element, and upload your reference set to it. From then on you can @-mention that Element directly in the prompt composer to pull the character into any shot.
4. Name the character in your prompt the same way every time. Pick one phrase — "the woman in the red raincoat" — and reuse it verbatim across shots. Consistent language helps the model stay locked on.
5. Generate one shot at a time. For each new shot, write the prompt as if continuing a storyboard: same character description, new action, new camera.
6. Check the face at full screen before accepting a clip. Consistency failures are subtle at thumbnail size.

Tips and common mistakes

• Fix the outfit in words too. Reference images do most of the work, but stating "same red raincoat, same short black hair" in each prompt is cheap insurance.
• Extreme angles are where consistency breaks. Top-down shots, heavy backlighting, and faces at the edge of frame are the usual failure points. If a shot needs a weird angle, generate a few extra variations.
• Consistency degrades across long chains. If you're building a 10-shot sequence, re-check shot 8 against shot 1, not just against shot 7. Small drift compounds.
• Two characters in one shot is harder than one. Give each character distinct references and distinct names in the prompt ("the woman in red," "the man with the umbrella") so the model doesn't blend their features.
• Reuse your reference set. Once a character works, save those images. They're an asset — every future video in the series starts from the same set.

Want ready-made wording for these prompts? Our Seedance prompt library has copy-paste examples for character shots, transitions, and camera moves.

Lock composition and camera movement with clay render

Clay render control (also called greybox) is the advanced option: you upload an untextured 3D blockout — plain grey shapes, no materials — and Seedance 2.5 uses it to lock the composition, camera position, and camera movement of the shot. The model then renders the actual scene on top of that structure.

Think of it as handing the model a storyboard it can't argue with. Where text prompts describe what you want and reference images show it, a clay render defines it geometrically.

When to use it

This is overkill for casual clips. Reach for it when:

• You need a specific camera move — a precise orbit, a dolly through a doorway, a crane up over a scene — and prompt wording keeps getting you close-but-not-quite results.
• Composition must match a layout, e.g. leaving clean space for text overlays, or matching a brand's storyboard exactly.
• You're recreating a shot across multiple videos and need the framing identical every time.

Step by step

1. Build or export a simple blockout. Any 3D tool works — rough boxes and cylinders are fine. Nobody sees this; it only needs correct proportions, positions, and camera.
2. Render or screenshot the clay view from the camera angle(s) you want, matching your output aspect ratio.
3. Upload the clay render as a structural reference alongside your other references.
4. Write a prompt for the look, not the structure. The clay render owns composition and camera; your text should describe surface details — "a sunlit Mediterranean alley, laundry lines overhead, warm afternoon light" — not "camera orbits left," which the blockout already dictates.
5. Generate, then adjust the blockout if framing is off. With clay render, composition problems are fixed in the 3D file, not in the prompt.

Tips and common mistakes

• Don't fight the blockout with the prompt. Conflicting instructions ("low angle shot" over a blockout built at eye level) produce mush. One source of truth per aspect.
• Start simple. A single room or street with one camera move teaches you how literally the model follows the geometry. Complexity can come later.
• Combine with character references for full control: clay render owns the stage and camera, reference images own the actor, text owns the lighting and style.

Chain 30-second clips into longer videos

Each generation caps at 30 seconds, but Seedance 2.5 is built for multi-round extension. The reliable pattern:

1. Generate shot 1 with the techniques above.
2. Take its final frame — or use first & last frame to design that final frame deliberately.
3. Feed that frame as the first frame of shot 2, with your character references still attached.
4. Repeat, then stitch the clips in any editor.

Because each clip shares a frame with its neighbor, cuts land cleanly and characters stay stable — this is where first & last frame and character consistency pay off together.

Fix one moment without redoing everything

Two more features worth knowing once you're producing regularly:

• Timestamp editing lets you target a specific second-range of a clip and regenerate just that segment. When shot 3 is perfect except for two seconds of weird hand motion, you fix the two seconds instead of re-rolling the whole shot.
• Green screen editing and camera angle editing let you swap backgrounds or adjust viewpoint after generation — useful for repurposing one clip across formats.

Where to start

If you only take one thing from this guide: stop writing longer prompts and start using first & last frame. Pick two images with one clear transformation between them, describe the motion in one sentence, and generate. Once that clicks, add character references for your second shot, and you'll have the core Seedance 2.5 workflow — everything else is refinement.

Ready to write your first prompts? Grab starter examples from our Seedance prompt library, follow the eight-step tutorial for the full workflow, or head back to the homepage.