Use your character in Hailuo
Hailuo is the strongest of the five at animating a single still with large, physically believable motion, which makes hero.png a reliable first frame. Identity across clips comes from either the dedicated subject-reference mode (one face image) or, on the newer MiniMax-H3 model, multiple reference images plus native audio.
Capabilities
- First frame (image-to-video)Yes
- Last frameYes
- Reference images9
- Native extendNo
- Max clip15 s at 768P or 1080P on Hailuo 2.3 (1080P is 6 s only); up to 2K on MiniMax-H3
- AudioYes
Which files to use
hero.pngFirst frame (first_frame_image on Hailuo 2.3 and Hailuo 02). Output follows the 2:3 aspect of the still.closeup.pngSubject reference (S2V-01 character image, or the first reference slot on MiniMax-H3). One clear single-person face.three_quarter.pngLast frame (Hailuo 02 first-and-last-frame mode) for a shot that ends turned toward camera, or an extra H3 reference.profile.pngExtra reference on MiniMax-H3, or the last frame for a turn-to-profile shot.back.pngLast frame for a walk-away shot, or an H3 reference if the action shows the back.contact_sheet.pngNot used; Hailuo reference slots expect one subject per image.character.jsonBible text at the top of every prompt; keep the negative wording out of the prompt and rely on the reference.
Steps
- Open hailuoai.video (or the Hailuo app), sign in and start a new video.
- For a shot that opens on the sheet pose, choose image-to-video and upload hero.png as the first frame. For a shot that should not open on the sheet, choose the subject (character) reference mode and upload closeup.png as the single-person reference.
- Paste the bible text from character.json, then describe the action and camera in one or two sentences. Keep one main character in the prompt.
- Choose the model and length: Hailuo 2.3 gives 6 or 10 seconds at 768P or 6 seconds at 1080P; MiniMax-H3 goes longer and adds audio where it is offered.
- Generate and check the face against closeup.png and the outfit against hero.png; if the outfit drifts, add the clothing line from the bible verbatim.
- To end on a known pose, use the first-and-last-frame mode with hero.png as the first frame and three_quarter.png, profile.png or back.png as the last frame.
- Chain scenes by exporting the last frame of the finished clip and using it as the next first frame with the same bible text.
Prompt template
{{bible}} The character in the first frame or reference image stays exactly the same: same face, hair, skin tone, outfit and build. {{action}} Smooth natural motion, single subject, consistent lighting. Note: first-frame mode and subject-reference mode are separate request types on the V1 API; pick one per clip.
Tips
- closeup.png is the reference to use for subject mode: the official guidance is a single person, unobstructed, in focus, no heavy filters.
- Image-to-video output follows the input aspect ratio (anything from 2:5 to 5:2 is accepted), so the 2:3 sheet works as is.
- Prompts that describe one subject, one scene and one action hold identity far better than crowded prompts.
- 1080P is limited to 6 seconds on Hailuo 2.3; use 768P when you need 10 seconds.
- For a turn-around shot, use first-and-last-frame mode with hero.png and back.png rather than asking the model to invent the back of the outfit.
- There is no extend endpoint; chain clips with the last frame and keep the bible text identical in every prompt.
Caveats
- MiniMax-H3 (announced by MiniMax on 31 July 2026: up to 2K, up to 15 seconds, native stereo audio) is served by the newer V2 video API. The field names reported for it (reference images, an end image) and the 9-image limit come from third-party pages and were not confirmed from the official reference.
- Hailuo 2.3 (MiniMax-Hailuo-2.3) is image-to-video only: 768P at 6 or 10 seconds, 1080P at 6 seconds, 24 fps, no audio. First-and-last-frame mode is documented for MiniMax-Hailuo-02, not 2.3.
- Subject reference on the V1 API is the S2V-01 model, which takes a single character image. MiniMax now describes it as a legacy model; new integrations are pointed at H3.
- Base64 data URLs for first_frame_image are reported by resellers; the official page examples use hosted URLs. Image limits reported by third parties: JPG, PNG or WebP under 20 MB, short side over 300 px, aspect between 2:5 and 5:2.
- The consumer Subject Reference feature was labelled beta when documented; its current name and placement in the web app may differ.
- Audio is only on MiniMax-H3; the capabilities row reflects the newest model, not Hailuo 2.3.
API notes
V1: POST https://api.minimax.io/v1/video_generation with model (MiniMax-Hailuo-2.3, MiniMax-Hailuo-02 or S2V-01), prompt, duration and resolution, plus first_frame_image (URL) and optionally last_frame_image for Hailuo 02, or a subject_reference array with type character and the image for S2V-01; it returns a task_id. Poll GET /v1/query/video_generation with the task_id until the status is success and a file_id is returned, then GET /v1/files/retrieve with that file_id for the download URL. V2 (MiniMax-H3) uses GET /v2/query/video_generation/{task_id} and returns a content URL directly; tasks are queryable for 7 days.