FL MiniMax H3 Temporal Reshot Assembler

The back half of the temporal reshot workflow. FL MiniMax H3 Temporal Reshot Assembler takes the sampled replacement latent from your Beat KSampler, decodes it, extracts the exact interval you selected in the planner's editor, resizes it back to the source's dimensions, and splices it in - overwriting only those frames. Everything outside the selection is untouched, and the source's audio, frame rate, metadata, alpha, and bit depth all ride through.

That "only those frames" guarantee is the whole point of this node, and it's worth understanding how strict it is. The sampler doesn't regenerate a separate clip and hope it blends - it denoises full spatial frames only across the H3 tokens that intersect your selection, re-injecting the source latent everywhere else at every step. The assembler then decodes the working window, crops out exactly the authored interval, and lays it into the original source. The result is a native ComfyUI video output you can drop straight into Save Video.

Inputs

Three inputs, no options to mis-set:

• shot_plan - the exact Temporal Reshot Planner output. Not a different plan, not a re-run of the planner - the same one you sampled from.
• latent - the sampled latent from FL MiniMax H3 Beat KSampler.
• vae - the H3 video VAE used to decode the reshot.
• video (output) - the spliced result, ready for Save Video or anything else that takes ComfyUI's video type.

The failure modes that are actually designed in

This node validates its inputs hard, and two errors are worth knowing before you hit them:

• Source changed after planning. The planner records a fingerprint of the source video, and the assembler checks it. If you moved, replaced, or re-encoded the source file after planning, it refuses with "source video changed after planning - run the planner and sampler again." That's not a bug, it's protecting you from a silent misalignment between the saved plan and the frames you think you're editing.
• Decoded too few frames. If the latent doesn't contain enough frames for the selection, it errors rather than truncating. Re-sample with the right plan.

The optional edge_blend_frames in the planner's settings softens the two temporal joins inside the selected range; it never changes a frame outside the selection, and the assembler rejects an edge blend longer than the selection itself.

The honest hardware note

The assembler materializes and normalizes the full source video in memory so it can resample frames to the 24 fps editing timeline. Long or high-resolution sources eat substantial system RAM - this is a "close the browser tabs" node, not a "casually rerun" one. And while the output runs at H3's fixed 24 fps, the source's other properties (audio, metadata, alpha, bit depth) are preserved.

Install is the shared pack story: ComfyUI Manager → FL MiniMax H3, or cd ComfyUI/custom_nodes && git clone https://github.com/filliptm/ComfyUI-FL-MiniMaxH3.git, then restart. No extra Python deps, no bundled weights. If you're coming from Fill Nodes, update Fill Nodes past the version that still registers these MiniMax node IDs first so you don't get duplicate node registrations.