MiniMax H3 Retake Stitch

Retake Mode is the Director's surgical edit, and it's genuinely clever about H3's first/last frame design. Load a base video, flip Retake Mode on in the toolbar, mark a range - and the Director regenerates only that range, anchored on the base video's own frames either side. The frame before the range becomes first_frame, the frame after becomes last_frame, which is exactly what H3's anchors are for: the new material meets the old on both cuts instead of gluing onto nothing.

MiniMax H3 Retake Stitch is the second half of that feature. The Director hands you a regenerated middle; this node splices it back into the base video - base head + retake + base tail, video and audio both, resampled onto H3's 24 fps output grid.

How it works

The whole node is driven by the Director's retake_info output - a JSON string carrying the base video path, the start frame and the generated length. That's deliberate: because the range comes from the same data that ran the render, it can never fall out of sync with what was actually generated. The node cover-crops the retake to the generated canvas, slices the head and tail from the base file, splices them together, and handles audio as stereo 44.1 kHz (resampled if needed, silence-padded when a slice comes up short).

The polite bit: if no retake ran - the Director wasn't in Retake Mode, or no base video was marked - the node passes the frames through unchanged rather than raising. It sits at the end of the graph, so failing there would throw away a finished render over a node that simply had nothing to do. You can leave it wired permanently.

The inputs and outputs

• retake_info - from the Director's retake_info output. Wire it, that's the contract.
• images - the decoded retake, straight from VAEDecode.
• audio (optional) - the retake's audio from VAEDecodeAudio. Leave it out and the base video's own audio is kept throughout.
• keep_base_audio - default off. Flip it on to ignore the generated audio and keep the base video's soundtrack across the whole result - handy when you only retook the picture.

Outputs are images (full-length video frames at 24 fps), audio, and fps, ready for CreateVideo:

base video → Director (Retake Mode) → sampler → VAEDecode ─┐
                                   → VAEDecodeAudio ─┐       ├→ Retake Stitch → CreateVideo
                                    retake_info ─────┴───────┘

Installing and gotchas

Same pack, same install as the Director - ComfyUI Manager search MiniMax H3 Director, or git clone, no pip deps. The only real dependency chain is H3's own: this is a Director-pack node, so it needs the checkpoint, encoder and VAEs the Director needs too.

Two things to expect. The retake's length snaps to H3's 17k+5 frame grid (5 s becomes 124 frames), so your marked range comes back on grid frames - the in-point is exact, the out-point is the nearest grid step, not necessarily your mouse position. And remember the Retake range is anchored on the base video's own frames, so the base video needs to be on the timeline (loaded as your retake base) for the anchors to exist at all. If you're getting a pass-through instead of a splice, that's the first thing to check - the node tells you in the console when it has nothing to do.