← All posts

Seedance 2.5: build the shot from your references

Bring pictures, movement and sound into one video brief, with native audio and clips up to 30 seconds.

A cobalt and ivory tram approaches a passenger waiting at a sunlit stop.

Sometimes the idea for a film already exists in pieces. There is a character photograph, footage with the right camera movement, and an audio cue that sets the pace. Seedance 2.5 brings those ingredients into the SUPERMODELS video creation tool, where a prompt can explain how they should work together.

Give every reference a role

ByteDance’s model generates video and sound together. It supports clips from 4 to 30 seconds, or an intelligent duration setting. A reference-based request can contain up to 30 images, ten video clips and ten audio files. Label them in the prompt with references such as [Image1] and [Video1], then explain what each should guide.

For a product concept, use a still photograph to establish the object and a short reference clip to suggest camera movement. Describe the material, light and action you want to see. Keep the first attempt focused: a single useful movement will tell you more than several unrelated ideas in the same shot.

Plan the sound with the picture

Native audio lets dialogue, effects and music form part of the same brief. Put spoken words in quotation marks and say who speaks. Audio can also be switched off. A product turning on a table might need only a soft room tone; a short scene with two people needs a clear exchange and enough time for it to happen.

Try: “A slow camera move towards a wind-up toy on a wooden desk. The toy takes three small steps and stops. We hear its spring unwind, then the quiet room. No music.” Start short, watch the result, and decide whether the next version needs a different movement or more time.

Choose frames or reference media

A first image and optional final image provide a separate workflow from mixed references. In SUPERMODELS, those frame inputs cannot be combined with reference media. The provider documents adaptive framing for first/last-frame runs, and adaptive framing with intelligent duration for video editing.

Reference videos affect the price. Review the estimate after attaching them, then check motion, continuity and sound in the result before extending the idea.

Open Seedance 2.5 to prepare a shot. Capabilities and workflow details were checked against the official Replicate documentation.

Cover image prompt

Editorial cover artwork generated with OpenAI image generation. The exact prompt is below.

Illustrate this everyday scene for the SUPERMODELS illustration collection: A cobalt and warm-ivory city tram pulling up at a quiet stop, seen from platform height. A single burnt-orange doorway and curving rails carry the composition; no signage or lettering.

Contemporary SUPERMODELS editorial gouache illustration with clean cut-paper shapes, fine subtle risograph grain, warm ivory, cobalt blue, moss green and burnt orange, tactile matte painted paper, soft directional light, generous negative space and an elegant playful composition. Limit the palette to those colors and charcoal shadows. Use broad simplified shapes and visible dry gouache texture, as in a printed editorial illustration. Avoid photorealism, 3D rendering, glossy materials, ornate fantasy, clutter, outlines, gradients and collage grids. One self-contained full-bleed landscape artwork, exact 16:9 aspect ratio, with the focal subject comfortably inside the middle 75 percent. No text, lettering, logo, frame, borders or watermark.

← All posts