IAMCCS-SuperNodes AU+IMG2VID Exec Planner

The planner that turns audio into a segmented video plan

LTX-2's superpower is audio-synced generation - the same model produces video and audio together, which makes it the go-to for lipsync and dialogue work. The pain is that you can't generate a ten-minute take. You generate 5-to-20-second segments, each with context from the last, and stitch them. Do that by hand with frame math and VAE nodes and you'll be debugging "one wrong value breaks three segments later" until you hate the whole hobby.

This node is the author's answer to that. It's a planning node, not a generator: you hand it audio and it works out the segment grid, duration, frame counts, and overlap for a full multi-segment audio-to-video run - then hands that plan to the sibling SuperNode stages in this pack (render, VAE, finalize). Same power as a hand-built graph, way less chance to break it.

How it decides the plan

The core inputs are audio, fps (default 25), and a segment length. Two planning modes exist: manual_segment_seconds uses your segment_seconds value directly (5 to 60), while explicit_preset_seconds snaps to a preset. The default overlap_frames of 9 is the classic LTX extension overlap - enough for the VideoVAE to see continuity without wasting frames.

The genuinely clever knob is audio_preprocess_mode. Default melband_vocals_duration_math runs the audio through MelBand RoFormer to isolate the vocal track, then uses vocal duration to size segments - great for dialogue where you don't want a cut mid-word. raw_audio_only skips that if you don't have MelBand installed. There's a matching audio_img2vid_backend / audio_img2vid_mode pair (default "modern pipeline", "single generation") and a route_mode that decides whether you target a segment count or a single duration.

Outputs matter more than inputs here: plan_payload is the JSON plan downstream stages consume, planned_duration_seconds, total_frames, segment_count, and recommended_overlap_frames are the numbers you'll want to sanity-check, and linx is the contract object that threads the whole SuperNode chain together.

Setting it up - and the dependency reality

The node itself ships in IAMCCS-nodes (Manager: search "IAMCCS", or clone the repo into custom_nodes). But the planner is honest about what it needs: the SuperNodes wrappers orchestrate existing ComfyUI/LTXV nodes (EmptyLTXVLatentVideo, LTXVConditioning, LTXVImgToVideoInplace, LTXVAudioVAEEncode, and friends) plus VideoHelperSuite and MTB. And if you want the vocal-aware planning, you need the MelBand RoFormer nodes and the MelBandRoformer_fp32.safetensors model.

The README's failure-symptom list is worth reading before you complain: missing dependency → the graph validates with missing-node errors or a SuperNode reports a helper is unavailable; duplicated IAMCCS-nodes folders → startup fails or routes register twice; stale saved workflows → dropdown values land in the wrong fields (recreate the node). This is a workflow-composition node for people who've already got an LTX-2 audio-video pipeline running, not a first project. When it's working, though, an entire segmented audio-video shoot plans in one click.