--- title: "How to keep the same face across AI images and videos" url: https://whatsthatmodel.com/guides/same-face-ai-images-videos/ last_updated: 2026-10-07 language: en --- # How to keep the same face across AI images and videos To keep the same face, start with an approved identity and reuse it through the controls each model actually supports. Soul ID is a trained identity layer in the Soul family; a video model usually receives an approved starting image or reference assets instead. Repeating a name or seed alone does not guarantee facial consistency across models or shots. Last updated 2026-10-07 · By What’s That Model? Affiliate link: we may earn a commission if you subscribe. Your model choice comes first. How links work Video: https://d28lhcrx5qdowv.cloudfront.net/media/explore/media/e49233cbd94028feaafc895fc1350b276278a66ab9672b5a9cf7045c2b2fd239-video.mp4 Video: https://static.higgsfield.ai/kling-3/features/feature-3.mp4 Existing originals and official model previews illustrate format, not results from this guide. Soul 2.0 · Source ↗ Seedance 2.5 · Source ↗ Kling 3.0 · Source ↗ ## On this page. ## Identity is different from style | Control | Use it for | Limit | | --- | --- | --- | | Soul ID | A consenting person’s trained identity in supported Soul workflows. | Do not assume a portable identity file or automatic support in unrelated models. | | Reference image / starting frame | Anchoring a scene before animation. | An image can still drift under large motion or changed viewpoint. | | Image reference set | Endpoint-supported identity and scene constraints. | Roles, counts and modes depend on the integration. | | Prompt / seed | Describing attributes or making a run repeatable where supported. | Neither establishes identity preservation across model families. | Sources: Soul and Soul ID help · Soul ID training and permission requirements · checked 2026-10-07 ## Build and approve the identity - Choose a fictional adult persona or a consenting real person; keep permission and reference provenance. - For a real person’s Soul ID, the official training help currently asks for at least 20 photos, up to 80. Follow its quality guidance and account controls. - For an invented persona, curate an approved sheet and use the product’s supported fictional-character/reference workflow; do not imply that a pile of synthetic portraits meets real-person training rules. - Freeze a front, three-quarter and side view, plus key wardrobe attributes. - Generate a small still set and reject drift before moving into video. Sources: Soul and Soul ID help · Soul ID training and permission requirements · checked 2026-10-07 ## Carry the approved frame into motion Use the dataset’s input_roles for the exact route. Kling 3.0 exposes starting/end frames in the curated integration; Seedance 2.5 exposes reference modes and edits. Veo has starting-image support in this dataset, while broader references are marked unverified here. For image-to-video, describe the action and camera instead of redesigning the face. Begin with one restrained action. Review the eyes, jaw, hairline, age and mouth frame by frame, including the end of the shot. A strong first frame can deteriorate later. Sources: Soul and Soul ID help · Soul ID training and permission requirements · checked 2026-10-07 ## Debug drift one variable at a time If the face changes, reduce the turn angle or camera motion, shorten the scene and improve the reference set. Keep lighting and wardrobe stable during diagnosis. Do not repeatedly add contradictory facial adjectives. Provider marketing may describe consistency more broadly than our verified endpoint data. This is a documented workflow and a review checklist, not a tested identity-retention benchmark. Sources: Soul and Soul ID help · Soul ID training and permission requirements · checked 2026-10-07 ## Sources and verification. Last updated 2026-10-07. Provider claims and endpoint records are dated evidence; workflow advice is editorial guidance. - Soul and Soul ID help Checked 2026-10-07 - Soul ID training and permission requirements Checked 2026-10-07 - Higgsfield MCP image catalog, exported 7 October 2026 Checked 2026-10-07 - Higgsfield: Soul 2.0 Checked 2026-10-07 - Higgsfield: public plan/credit table, Italy EUR Checked 2026-10-07 - Higgsfield MCP video catalog, exported 7 October 2026 Checked 2026-10-07 - ByteDance: Seedance 2.5 Checked 2026-10-07 - Kuaishou: Kling AI 3.0 release Checked 2026-10-07 ## Continue the workflow. ## Common questions. ### Does Soul ID work in every video model? Do not assume that. The official help describes Soul-family identity reuse; video routes need their supported frames/reference assets. Confirm the exact route. ### How many Soul ID photos should I use? The current official training help says 20 or more photos, up to 80, of one consenting person. Quality and variety matter; older pages may describe different requirements. ### Can I guarantee the same face in every shot? No. Inspect every shot and regenerate or edit failures. This workflow reduces avoidable drift but does not prove perfect retention. ## Start with one reviewed draft. Affiliate link: we may earn a commission if you subscribe. Your model choice comes first. How links work ## Relevant links and sources - [Home](https://whatsthatmodel.com/) - [Guides](https://whatsthatmodel.com/guides/) - [Try Soul 2.0 on Higgsfield ↗](https://whatsthatmodel.com/go/soul-2-0?pt=guide&pos=hero&v=default) - [How links work](https://whatsthatmodel.com/affiliate-disclosure/) - [Soul 2.0](https://whatsthatmodel.com/models/soul-2-0/) - [Seedance 2.5](https://whatsthatmodel.com/models/seedance-2-5/) - [Kling 3.0](https://whatsthatmodel.com/models/kling-3-0/) - [Soul 2.0 · Source ↗](https://whatsthatmodel.com/go/official-soul-2-0?pt=guide&pos=examples&v=media) - [Seedance 2.5 · Source ↗](https://whatsthatmodel.com/go/official-seedance-2-5?pt=guide&pos=examples&v=media) - [Kling 3.0 · Source ↗](https://whatsthatmodel.com/go/official-kling-3-0?pt=guide&pos=examples&v=media) - [Identity is different from style](#article-section-1) - [Build and approve the identity](#article-section-2) - [Carry the approved frame into motion](#article-section-3) - [Debug drift one variable at a time](#article-section-4) - [Soul and Soul ID help](https://whatsthatmodel.com/go/source-9992eb6e373f?pt=guide&pos=source&v=citation) - [Soul ID training and permission requirements](https://whatsthatmodel.com/go/source-fe3cc5cebb0a?pt=guide&pos=source&v=citation) - [Model controls ↗](https://whatsthatmodel.com/models/) - [Animate a photo ↗](https://whatsthatmodel.com/guides/photo-to-video-ai/) - [Higgsfield MCP image catalog, exported 7 October 2026](https://whatsthatmodel.com/datasets/higgsfield-image-catalog-2026-10-07.md) - [Higgsfield: Soul 2.0](https://whatsthatmodel.com/go/source-b629b6b09a31?pt=guide&pos=source&v=citation) - [Higgsfield: public plan/credit table, Italy EUR](https://whatsthatmodel.com/go/source-ba6feeb2bdda?pt=guide&pos=pricing-ref&v=citation) - [Higgsfield MCP video catalog, exported 7 October 2026](https://whatsthatmodel.com/datasets/higgsfield-video-catalog-2026-10-07.json) - [ByteDance: Seedance 2.5](https://seed.bytedance.com/en/seedance2_5) - [Kuaishou: Kling AI 3.0 release](https://ir.kuaishou.com/node/11216/pdf) - [character consistency ↗](https://whatsthatmodel.com/best/character-consistency/) - [ai influencers ↗](https://whatsthatmodel.com/trends/ai-influencers/) - [seedance 2 5 vs kling 3 0 ↗](https://whatsthatmodel.com/compare/seedance-2-5-vs-kling-3-0/) - [Try Soul 2.0 on Higgsfield ↗](https://whatsthatmodel.com/go/soul-2-0?pt=guide&pos=final&v=default)