MiniMax H3 Latent Extend (continue clip):
MiniMaxH3LatentExtend prepares a finished MiniMax H3 clip for continuation. It extends the clip’s underlying latent data along the time axis, keeps the existing portion pinned, and marks the new tail as available for generation. This lets the model continue the clip while seeing its earlier content at every sampling step. The node also accounts for the frame lengths H3 supports and can limit the extension when a reference is too short to cover the resulting clip.
MiniMax H3 Latent Extend (continue clip) Input Parameters:
samples
The finished MiniMax H3 latent clip you want to continue. This required LATENT input must contain a nested H3 latent with video data; an ordinary or incompatible latent will not work. The node uses its current duration to calculate the extension and preserves the existing clip as the pinned beginning of the result. No minimum, maximum, or default is specified for this connection.
extra_frames
The number of rendered frames you want to add. This required INT input defaults to 34, with a minimum of 1 and a maximum of 3600, in steps of 1. H3 uses supported frame lengths on a 17k+5 grid, so the node rounds the total length up to a supported value. The actual number of added frames may therefore be slightly larger than you requested. A reference-length cap can reduce the result instead.
strength
Controls the strength applied when the node prepares the newly added region for generation. This required FLOAT input defaults to 1.0, ranges from 0.0 to 1.0, and changes in steps of 0.05. At 1.0, the new tail is built from noise. Because that region starts empty, lower values have no existing tail content to pull back toward; keep the default unless you are deliberately experimenting with the effect.
feather_latent_frames
Controls how softly the transition between the existing clip and the new tail is masked. This required FLOAT input defaults to 0.5, ranges from 0.0 to 8.0, and changes in steps of 0.5. The value is measured in latent frames rather than rendered frames; one latent frame is about three rendered frames. Increasing it softens the join across more of the boundary, while 0.0 requests no feathering.
reference
An optional VIDEO reference for the run. When connected, the node checks how many frames it contains and caps the extension if necessary, because the reference must cover the entire resulting clip, not just the new tail. This can expose a length problem here instead of leaving it for a later preparation node. There is no default video or numeric range; leave it disconnected when you do not need a video-reference cap.
reference_frames
An optional, manually entered reference-length cap for cases where the reference is not connected as VIDEO. This INT input defaults to 0, has a minimum of 0 and a maximum of 100000, and uses 1-frame increments. A value of 0 means no manually entered cap. When a VIDEO reference and a nonzero manual cap are both supplied, the node uses the smaller available frame count. The node reserves a frame and rounds the usable limit down to a supported H3 length, so the final length can be below the number entered.
MiniMax H3 Latent Extend (continue clip) Output Parameters:
samples
The extended LATENT result. It retains the original clip’s latent content, adds space for the continuation, and includes a noise mask that keeps the existing portion pinned while making the new tail available for generation. Connect this to the subsequent H3 generation workflow rather than treating it as an already rendered extended video.
length
The final total length in rendered frames, returned as an INT. This is the value to use for downstream conditioning: it reflects H3’s supported frame-length adjustment and any reference cap, so it may differ from the original clip length plus the requested extension.
report
A STRING summary of what the node prepared. It gives the original and final frame counts and latent time lengths, the number of frames added, the number of initial frames pinned, and the conditioning length to set. If a reference shortened the extension, it also reports the cap and the requested length. The same report is displayed in the node’s UI.
MiniMax H3 Latent Extend (continue clip) Usage Tips:
- Connect a finished MiniMax H3 latent to
samples, setextra_framesto the continuation you want, and checkreportfor the actual added duration after H3 length adjustment. - Feed the
lengthoutput into the conditioning node. Using the source clip’s old length will no longer match the extended latent. - Start with
strengthat its default of 1.0. Lower it only when testing how a less strongly prepared new tail behaves. - Start with
feather_latent_framesat 0.5. Adjust it when evaluating the join, remembering that this setting counts latent frames, not rendered frames. - Connect
referencewhen your run uses a VIDEO reference, or setreference_frameswhen you need to enter its frame count manually. Check the reported final length: the reference must cover the whole clip, including the original portion.
MiniMax H3 Latent Extend (continue clip) Common Errors and Solutions:
expected a MiniMax H3 nested latent, got <type>
- Explanation: The value connected to
samplesis not the nested latent format this node expects. The displayed type name depends on what was supplied. - Solution: Connect the finished latent from a compatible MiniMax H3 workflow, rather than a generic latent or another data type.
no video tensor in this latent
- Explanation: The supplied latent has the expected nested structure, but the node cannot find video data in it to extend.
- Solution: Check the node feeding
samplesand use a MiniMax H3 latent that contains the finished video clip.
nothing to extend: the reference allows at most <length> frames and the clip is already <old_frames>. Supply a longer reference, or disconnect it.
- Explanation: After accounting for the reference limit and supported H3 lengths, the allowed final duration is no longer than the existing clip.
- Solution: Supply a reference that covers a longer total clip, increase an overly restrictive
reference_framesvalue, or disconnect the reference cap if the run does not need it.
Downstream conditioning length does not match the extended clip
- Explanation: The node’s description and report warn that the extended clip’s length differs from the source length. A downstream conditioning node set to the old length will not match the prepared latent.
- Solution: Connect the
lengthoutput to conditioning, or set conditioning to the exact final length shown inreport.
capped at <length>: the reference holds <frames> frames and has to cover the whole length. Asked for <requested>.
- Explanation: This is a report notice, not an error. The requested extension exceeded the length the supplied reference could support, so the node shortened the result.
- Solution: Use the reported length if it meets your needs. For a longer continuation, supply a longer reference or correct an unnecessarily low manual
reference_framescap.
