Seedance 2.5 from scratch: character sheets, prompts with Claude and 30 seconds without a cut

I asked Seedance 2.5 for two mariachis serenading a balcony on a colonial Mexican street at night, in the rain, film-noir style. It gave me that — plus a story: she leans out, smiles, the other mariachi sees it, gets angry and attacks him. It closed the tale on its own. That’s the best and the most delicate thing about the model: if you don’t tell it, it assumes and executes. This is the guide to how I work it so that I’m the one directing, not the model.

The serenade: characters from turnaround sheets, environment generated in GPT Image 2.

The facts

• Each generation takes up to 30 assets — characters, props, locations — and runs up to 30 seconds in a single take, no stitching.
• You can raise the bitrate, and the model makes targeted corrections on specific objects or regions of the frame.
• A 30-second generation took me around 6 minutes; I saw no speed difference between the unlimited-plan queue and the paid one.
• If a generation gets rejected — it happened to me for “copyright restrictions” — Higgsfield refunds the credits.
• Higgsfield runs promos constantly; at the time, 33 days unlimited of 2.5 in 15-second clips at 720, one at a time.

Step 1: characters before video

I never start in video. I start in image, with a turnaround character sheet, white background, no text. I make it in GPT Image 2 — I liked its style better — though Nano Banana works too. Out of that come the mariachi in the white suit, the stronger one in the black suit (the bad guy) and the girlfriend. For the anime, I give GPT Image 2 my original reference and ask for “this character in anime style”: it returns several versions and I pick. Same with the environment: one image, then to video with the references tagged as @image 1, @image 2.

Step 2: a prompt system, not loose prompts

This is the line between playing and producing. I train a project in Claude with the visual style, the kinds of shots we use, duration, genre and the reference images connected. I have the Higgsfield MCP hooked up, so Claude knows how to structure a prompt for Seedance. From there, every new shot comes out consistent with the continuity of the scene. You can do it inside Higgsfield; I prefer my systems because the day I dropped a letter — I typed “bacon” instead of “balcón” — the model had no way to correct me.

The project in Claude: style, shots, duration and references; every prompt comes out of it.

Step 3: shot by shot, with extend video

A 30-second shot exists, but almost everything lives better at 15. What I do is generate and, if I like it, extend video with the previous shot as reference: that’s how I go shot by shot building the scene, and the sequence — the group of scenes — comes from adding extensions. You can feed audio references: your voice, your music, and there’s a voice-change function for when “it sounds Spanish but they’re not saying anything,” which happens.

Desktop is where it works best: grab your video, modify blocks, move them, add a shot, draw what to erase and what to add, change resolution and durations. That’s for whoever wants something more detailed; out of the gate, with an image and a prompt you already have a scene.

What the model doesn’t hold you back on

I’m doing effects for a horror film: there’s gore, there’s blood. Other models limit me; on Higgsfield I haven’t found the model holding back the violence. It also depends on how you describe things. And what I don’t use for generating full content, I use for what I actually need: set extensions, VFX that used to be too ambitious, plus educational videos and infographics — a cartoon project with a 3D armadillo came out of here.

The cartoon characters: you build in blocks, scene by scene.

The LATAM read

I saw a lot of fully AI-generated advertising on planes and on TV during the World Cup. If you’re an agency, this already has to be in your catalog of skills. For the independent filmmaker here, the read is different: when you don’t have the crew, the budget or the time to bring a story to life, this is a very good tool for telling it. Lower resolution is cheaper: I make the project on the version that exists and when the new one comes out I scale it, without rushing and spending less.

My personal opinion — Daniel Blanco A.M.C.

It’s so good I don’t know what to think or feel. The movement continuity is there, there’s a directing intention created by the model, and that has consequences — which is why there’s so much hate toward this technology: it touches the identity of artists. I think it’s an instrument to shape. I never think about the model or the specific result, but about what it’s going to become later. And to date, no video generation model has caught up to Seedance. If you’re going to get into this market, today is the best day.

Your first scene

Make one character sheet, one environment image and one 15-second prompt. Then extend it once. With two shots you know whether the workflow is yours — and if the model over-directed you, you know what you need to nail down.

Images: frames from Daniel Blanco A.M.C.’s video.

Original source: Seedance 2.5: complete guide from scratch, on YouTube · Seedance 2.5: 30 seconds, one take

Extra for you: if you’re experimenting with AI for video, my Higgsfield link has discounts and promos to get started — and using it supports this content.