Download for Mac

Guides

How to make AI video that holds up.

A good take comes from what you prepare before you press Generate: a character sheet the model can hold, the right references in the right order and a prompt that names one shot at a time. These guides are what we learned making videos with Seedance, Kling and Higgsfield, with the failures and what they cost.

Characters.

A character the model keeps from shot to shot: the reference sheet, full body and a close face, the same light and clothes.

A real character reference sheet for an original fictional woman: a face close-up, then full body front, three-quarter and back, on one warm studio backdrop in the same green field jacket
HiggsfieldAI Character Reference Sheet: One Face That HoldsHow to build a character sheet an AI video model keeps across shots: one face per sheet, full body plus a close face, the angles, clothes and checks.
Two people at a microphone swapped for two new characters who keep the same moves and timing
Character swapHow to Swap a Character in a Video With AIFour ways to replace a person in a clip with a new character, a step-by-step swap on a Mac, what makes a swap hold, and what a 10-second swap costs.
A reference board with a headshot and a full-body image for each character, 1 to 8 subjects per request
SeedanceSeedance Character Consistency: References That HoldHow Seedance 2.0 and 2.5 keep one character across shots: reference limits, headshot and full-body images, character sheets, prompt wording, what breaks it.

References.

Which images and clips to give a model, how many, in what order, and when a first or last frame beats a reference.

Face width in a 480p Genjutsu draft: from 130 px the eyes hold, 80 to 130 px they go soft, under 80 px they break
GenjutsuHow to Use Genjutsu: Sources, References and PromptsA Genjutsu swap step by step: which source clips work, how big a face must be, upload order, a prompt template, Lite vs full, and fixes for the usual faults.

Prompting.

How to write what the camera sees: shot size, one move per shot, timing for multi-shot clips, dialogue and sound.

The Seedance prompt formula as chips: subject, action, scene, lighting, camera, style, quality, constraints
SeedanceSeedance 2.5 Prompt Examples: Multi-Shot, Dialogue, 30 s9 ready Seedance 2.5 prompt examples, from a multi-shot montage to a 30-second one-take and a two-voice dialogue, with the structure ByteDance's guides use.

Fixing takes.

Read a failed take before you pay for another: what went wrong, and the one change that fixes it.

A repair flow for a bad Higgsfield take: rejected, failed or wrong, then rerun once and change one thing at a time
HiggsfieldFix a Bad Higgsfield Take: Diagnose Before You RerollWhy a Higgsfield video was rejected, failed or came out wrong, and how to fix each case: moderation codes, one change at a time, layout maps and references.

Workflow and cost.

Cheap drafts first, the final render once: the order of work that keeps a video from costing twice.

The Higgsfield workflow in four steps: stills first, a cheap draft, a check of the frames, then one final render
HiggsfieldHow to Generate on Higgsfield Without Wasting CreditsA Higgsfield workflow that works: pick the model by task, approve stills before video, draft cheap, write prompts in blocks, check a take before paying again.
Six numbers for Seedance 2.5: 30 seconds, 30 reference images, 10 video clips, 10 audio clips, 1080p and $0.46 a second
SeedanceSeedance 2.5 Guide: Limits, Modes and Price per SecondSeedance 2.5 makes 30 s of video with sound from up to 50 references and costs $0.10 to $1.14 a second by API. Its modes, limits and price on each channel.

Try it from your Mac.

Jutsu Studio is a free, open-source Higgsfield desktop app for Mac: Seedance, Kling, Wan, Genjutsu and the rest of its models.

  • The price is on the button before a job runs
  • A queue, cheap drafts first and a daily spending cap
  • Every result kept in a local library
Download for MacFree · macOS 14+ · v0.2.6 · 25 MBNo Higgsfield account? Sign up

FAQ.

Where do these guides come from?

From the videos we make with Higgsfield, Seedance and the other models while building Jutsu Studio: what failed, what fixed it and what it cost.

Which models do the guides cover?

Mostly Seedance 2.5, Higgsfield Genjutsu and Wan, with Kling where it fits: the models people run on Higgsfield and through video APIs. Each guide shows the date it was last updated.

How do I keep a character consistent across AI video shots?

Make one reference sheet per character before any video: a full-body shot and a close face, front and three-quarter, on a plain background in even light, in the clothes the character wears in every scene. Give the model the same sheet in every shot and describe only what changes.