Continuity and chaining#
A take is one render of one shoot. Anything longer than a take, and anything that must look the same from one take to the next, comes down to what the model is anchored to. This page lists the three anchoring modes of the video models, the two ways Maystro chains renders, and how a reference shoot chains.
The three anchoring modes#
| Mode | You give | What is anchored | Framing | Voice references | Chain clips |
|---|---|---|---|---|---|
| Text-to-video | the prompt only | nothing beyond the prompt; identity comes from the prompt and the LoRAs | free | no | yes |
| Image-to-video | a start frame | the frame, hard: the video continues from it, geometry and identity come from it | fixed by the frame | no | yes |
| Reference-to-video | 1 to 9 reference images and up to 3 voice clips | identity, look and setting, softly; the voice timbre | free: references do not fix the framing — unless the shoot continues from the previous shoot's last frame | yes, clip 1 (every clip with Keep the references) | yes: clip 1 from the references, later clips from the last frame, optionally with the references kept |
LTX presets are image-to-video. MiniMax H3 presets are text- or image-to-video; the H3 reference presets, and every character set preset, are reference-to-video. The two H3 kinds are different model variants, and a shoot runs in one mode: a reference shoot has no start frame of its own — it can only continue from the previous shoot's last frame.
Two kinds of chaining#
Generate all shoots (projects, scenes, shoots) renders the shoots of a scene one after another, each as its own take. The last frame of a take is handed to the next shoot as its start frame. An image-to-video shoot continues from that frame. A reference shoot ignores it unless Continue from the previous shoot's last frame is ticked on it: its identity and setting come from the references and the scene constants, so the cut between two reference shoots is a real cut. The handover crosses scene boundaries: a scene's first shoot continues from the last shoot of the previous scene, in Generate all shoots as in Generate whole project and in a single Generate.
Chain clips render one shoot as several clips inside one take. The main prompt is clip 1; every continuation clip you add is one more clip, rendered image-to-video from the last frame of the previous clip, and the clips are joined automatically. This is how a take gets longer than the 362-frame ceiling, and it is faster than one long take of the same length.
Chain clips#
Chain clips are available on every MiniMax H3 shoot, reference shoots included (see below). In the shoot panel, Add continuation clip adds a clip; a take holds at most sixteen clips (five until October 2026). A long chain simply takes as long as its clips do: the wait budget grows with the clip count, and cancelling still takes effect at the next clip boundary.
- Write only the beat of the clip in each box. The Scene constants field is added to every clip, so the setting, the characters and the light are restated without retyping.
- Every clip renders the shoot's frame count; the joined take is
clips × frames − (clips − 1)frames long. Two clips of 124 frames give 247 frames, about 10 seconds. - The seed advances by one per clip.
- The join is a hard cut with no visible transition: the first frame of each continuation clip duplicates the last frame of the previous clip and is dropped, and the audio crosses over one frame so the seam does not click.
- Progress runs across all clips as one bar. Cancelling takes effect at the next clip boundary.
The background can still drift over several clips. Restating the setting in the scene constants is the fix; see Prompting.
Reference shoots and chaining#
Chain clips work on shoots with reference images, and so on character set presets. Clip 1 is rendered reference-to-video from the references. Every continuation clip continues image-to-video from the last frame of the previous clip, without the references: the reference layout has no start frame, so the two cannot meet in one clip. The shoot panel says so next to the chain controls.
What that means for a reference chain:
- Identity reaches the continuation clips through the frame they start from, and through the character LoRA of a character set. As in any image-to-video chain it can drift over several clips; a character set keeps it steady, and a short chain drifts less than a long one.
- Voice references anchor clip 1 only. The continuation clips carry the voice of clip 1 the way they carry its picture: usually close, not guaranteed.
- Prompts: the picture numbering of the references does not exist in the continuation clips. Describe the character in words in every clip; the Scene constants field does that for you.
Keep the references in the continuation clips changes that trade. With the box ticked, every continuation clip is rendered from the last frame of the previous clip and the references, so the setting cannot be re-imagined when the camera reveals a new part of the room, and the voice references anchor every clip, not only clip 1. The frame anchor stays as hard as before. The cost is time: each continuation clip carries the reference images through the model again and takes roughly 2.4 times longer than a plain continuation clip, about 3 times with a voice reference on top. Use it for scenes where the camera moves through the room or where the voice must stay locked; leave it off for a static framing where a plain chain is already stable.
Keep only the voice references in the continuation clips is the cheap middle ground for a shoot with voice references. Every continuation clip is rendered from the last frame of the previous clip and the voice references, without the reference images: the voice stays anchored along the chain while the setting still comes from the frame alone, and a continuation clip takes only about a tenth longer than a plain one. Without it, the later clips of a plain chain can drift in timbre and level and even garble a spoken line, so on a shoot with voice references this box is the one to tick. The box appears only on a shoot with voice references and a runtime that supports it, and it is greyed out while Keep the references is ticked, since that already carries the voice. On a shoot that newly gets voice references and chain clips the box comes ticked; untick it and the choice sticks for that shoot. Shoots from before this option keep their old behaviour until you tick it.
On an older runtime the reference model cannot render image-to-video; the shoot panel then disables Add continuation clip on reference shoots and says why. The Keep the references box appears only when the runtime supports it. Update Maystro to chain reference shoots.
For a longer reference scene, splitting the action into shoots and using Generate all shoots is the other option: every shoot then gets the references and the voice references, and with the option below the shoots follow each other like the clips of a chain.
Continue from the previous shoot#
Continue from the previous shoot's last frame sits under the references of a reference shoot. Ticked, the shoot is rendered from the previous shoot's last frame and its own references and voice references — the same keyframe plus reference layout as a chained continuation clip with Keep the references — so the shot picks up where the previous take ended while identity, setting and voice stay anchored. Unticked, the shoot is rendered from its references alone and the cut is a real cut, as before.
- The previous shoot is the one before in project order: a scene's first shoot continues from the last shoot of the previous scene.
- The frame is the previous shoot's selected take's last frame; without a selection, its latest completed take's. The panel shows the frame it will use, or says that the previous shoot has no completed take yet.
- Generate all shoots and Generate whole project hand the frame down the chain the moment the previous take completes; a single Generate reads it from the previous shoot's takes at that moment.
- The shoot can still chain clips: clip 1 continues from the handed frame with the references, the continuation clips follow as the chain options say.
- Cost: the frame rides next to the references, so the take costs about what a continuation clip with Keep the references costs — a few percent more than the references alone on the default reference canvas.
The box appears only on a runtime that renders a start frame next to the references; on an older runtime update Maystro.