Skip to content

How to Keep Your AI Character Consistent Across Every Generation: Soul ID Explained

2026-06-29 · blog how-to · Higgsfield · original

Topic: Assets, references & consistency (also: Post-production & finishing, Video generation & shot direction, Failure modes & fixes) · mode: Explainer · tags: vendor claims

This Higgsfield blog explainer presents Soul ID, a face trained once from 20+ clear photos for 25 credits (about $1.25), as the main way to keep an AI character consistent. Every clip is then anchored with a Soul ID start frame, an Element reference, the same written description and last-frame chaining, starting from an image rather than text and chaining only two or three clips before a fresh portrait. Drift in a finished clip is fixed with Genjutsu Object Swap.

What is said

  • Summary and FAQ

What is said

★ — how much it matters to the source.

  • ★★ Higgsfield's blog says an AI character changes faces between shots because most generators rebuild the person from the prompt on every run with no memory of earlier results. · character drift
  • ★★ Soul ID trains a digital avatar from your own photos, or from someone else's when you hold usage rights for them. · Soul ID
  • ★★ Soul ID training accepts up to 80 reference photos, with 20 or more recommended. · Soul ID
  • ★★ For the best Soul ID quality each reference photo should be 960 px or higher. · Soul ID
  • ★★ Training a Soul ID takes 3 to 5 minutes. · Soul ID
  • ★★★ Creating a Soul ID costs 25 credits, about $1.25, at Higgsfield's standard credit rates as of October 2026. · Soul ID
  • ★★ Soul ID runs in Soul 2.0 and Soul Cinema Preview, on any paid Higgsfield plan. · Soul ID, Soul 2.0, Soul Cinema
  • ★★ Higgsfield says a trained Soul ID is stable across sessions, repeatable across different models and usable for stills, video clips and full sequences. · Soul ID
  • ★★ Higgsfield's blog says face drift grows with content volume: at scale, inconsistency stops being a minor flaw and becomes the first thing viewers notice. · character drift
  • ★★ The blog's Soul ID steps: upload about 10 or more clear photos with varied angles and expressions, train the avatar, pick a style preset, generate, then reuse it. · Soul ID
    • a number in question (doubtful): "Around 10 or more" — check against the recording
  • ★★ Soul ID offers 50+ curated style presets, such as Amalfi Summer, Gorpcore Outdoor and 0.5 Selfie. · Soul ID
    • a number in question (doubtful): "50+ curated styles" — check against the recording
  • ★★ Higgsfield calls Soul ID the main method for holding a character, as the only one that stores the identity itself; four other techniques can be used alone or on top of it. · Soul ID
  • ★★ In Seedance on Higgsfield, an @character reference holds face geometry, skin tone and style within one clip. · Seedance
  • ★★ For the same character across many separate clips, Higgsfield advises saving it as an Element (click @ Elements or type @, pick from My Elements); Soul ID characters appear there automatically. · Elements, Soul ID
  • ★★ Cinema Studio takes up to 9 references in version 3.0 and up to 50 in version 4.0; its Elements are shared across shots and can be pulled from other projects. · Cinema Studio 4.0, Elements, Cinema Studio
  • ★★★ Higgsfield advises starting a clip from an image rather than text: image-to-video holds a character better because the first frame already has the face, outfit and framing. · image-to-video
  • ★★ Generate the portrait in Soul 2.0 or Soul Cinema Preview with the Soul ID selected, upload it as the start frame in Seedance or another video model, and keep the prompt to action and camera move. · Soul 2.0, Soul Cinema, Soul ID, Seedance, keyframe
  • ★★ Higgsfield advises writing one character description (hair, clothing, distinguishing features, age range) and pasting it unchanged at the start of every prompt; text cannot pin a face but stops reinterpretation.
  • ★★ In Cinema Studio the character description goes in the Project brief, which applies across the whole project and is read by the AI Director when it drafts shots. · Cinema Studio
  • ★★ Higgsfield advises using the last frame of one clip as the first frame of the next; in Seedance, generate both frames as references and use first-and-last-frame for the transition. · Seedance, keyframe · also in: How do I use Seedance?
  • ★★ In Cinema Studio 4.0, Forward / Backward Extend takes an uploaded video of up to 30 seconds and continues it in either direction with the character still on screen. · Cinema Studio 4.0, AI Video Extender
  • ★★★ Since small errors accumulate over long chains, Higgsfield advises chaining only two or three clips from one start frame, then starting a new chain from a fresh Soul ID portrait. · Soul ID, keyframe
  • ★★ If a face drifted in a finished clip, Higgsfield advises Genjutsu Object Swap: upload the clip with reference photos of the character, and it is corrected without regenerating the scene. · Genjutsu, Object Swap
  • ★★ Genjutsu Motion Transfer takes only the movement from a video and builds a new scene around it, so one character can repeat a performance across clips with motion, camera and timing intact. · Genjutsu, motion transfer
  • ★ Restyle, inside Genjutsu Motion Transfer, redraws the whole video in one of 20+ animation styles while the motion stays as filmed. · Genjutsu, motion transfer
  • ★★★ Genjutsu takes one source video of 4 to 30 seconds and up to 30 reference photos per generation, outputs 480p, 720p or 1080p, and offers 30+ presets, so a prompt is optional. · Genjutsu
  • ★★ Genjutsu path: Video → Higgsfield Genjutsu, then upload the source clip and references for what should change, one for a simple swap and more for a full rebuild. · Genjutsu
  • ★★ Higgsfield suggests Soul ID portraits as Genjutsu's reference photos, linking the trained identity to existing footage. · Genjutsu, Soul ID
  • ★★★ For Soul ID training Higgsfield advises 20+ clear photos without sunglasses, masks or extreme expressions, and consistent lighting across them so lighting inference is accurate. · Soul ID
  • ★★ Higgsfield advises picking Soul ID style presets that reflect the brand identity early, since switching too much later may confuse audience memory. · Soul ID
  • ★★ Higgsfield says reference matching anchors to one photo's angle and lighting and drifts when the scene changes, while Soul ID training builds an internalized model of the face. · Soul ID
  • ★★ Higgsfield says a dolly, orbit or tracking shot defined in the prompt is executed by Cinema Studio across every model on Higgsfield, including Seedance 2.0, Veo 3.1, Kling 3.0 and Wan 2.6. · Cinema Studio, Seedance 2.0, Veo 3.1, Kling 3.0, Wan 2.6
  • ★★ Higgsfield suggests one spokesperson trained in Soul ID can appear in a Seedance 2.0 commercial, a Kling 3.0 narrative sequence and a Veo 3.1 close-up with the same face and planned camera moves. · Soul ID, Cinema Studio, Seedance 2.0, Kling 3.0, Veo 3.1
  • ★ Higgsfield advises updating the Soul ID avatar periodically if the brand evolves, while keeping the core identity stable for audience recognition. · Soul ID
  • ★★★ Higgsfield's recipe: train the identity once in Soul ID, then anchor every clip with a Soul ID start frame, an Element reference, the same written description and last-frame chaining; fix drift with Genjutsu Object Swap. · Soul ID, image-to-video, Elements, Seedance, Cinema Studio, Genjutsu, Object Swap
  • ★ For brand campaigns, Higgsfield's blog says twenty variations of the same spokesperson reinforce the message twenty times, while twenty slightly different faces reinforce nothing. · Soul ID
  • ★★ Higgsfield's FAQ says a trained Soul ID applies automatically to both image and video generation on Higgsfield, including Seedance 2.0, Veo 3.1 and Kling 3.0. · Soul ID, Seedance 2.0, Veo 3.1, Kling 3.0, image-to-video

Threads in this video

Mentioned

Central: Cinema Studio · Genjutsu · Soul ID

Also: Concepts: character drift · image-to-video · keyframe · motion transfer

Products: AI Video Extender · Cinema Studio 3.0 · Cinema Studio 4.0 · Elements · Kling 3.0 · Object Swap · Restyle · Seedance · Seedance 2.0 · Soul 2.0 · Soul Cinema · Veo 3.1 · Wan 2.6

Replies to

← Previous: How to Make AI Video Look Like It Was Actually Filmed: WAN Camera Control Guide · Next → This AI VFX is Unreal! 🤯