H3RelayInternalLTXRollingInput

Every LTX enhancement of an H3 shot needs a window of pixels to start from: the current segment plus enough of the previous shot's tail to keep the diffusion temporally coherent. Rolling Input is the node that assembles that window. It's the first of the four LTXRolling* nodes in the enhance graph, and its job is to turn disk paths into the pixel stack the rest of the machinery encodes.

The description says it cleanly: Build one incremental LTX enhancement window from the current delivered H3 segment and the exact tail stored in the previous accepted checkpoint.

The inputs

• current_segment - a relative or absolute path to the current delivered H3 segment inside ComfyUI output. In the expanded graph this is wired automatically from the raw shot's disk record.
• previous_checkpoint - for shot 2+, connect or enter the previous accepted H3 checkpoint. Its exact decoded tail becomes LTX context. Leave blank for the first shot. This is the pixel-side counterpart of the latent tail that Rolling Inject freezes later - the input window needs the previous shot's actual pixels too.
• context_frames (default 17) - pixel frames carried into LTX. The tooltip gives the exact conversion: seventeen maps exactly to three LTX temporal latents. That's why 17 is the default - it's the smallest context that lines up with LTX's latent grid.

The outputs

• padded_images - the assembled window: current segment plus context tail, ready for VAE encode. This is what feeds the encode and later the IC-LoRA guide.
• rolling_context - the H3_RELAY_LTX_ROLLING_CONTEXT token. It records original_frames, context_frames, and delivered_frames, and it's threaded to both RollingCrop (which needs those counts to trim padding) and AcceptLTX.
• original_frames / delivered_frames - the same counts surfaced as plain ints for inspection.
• status - text summary of what the window contains.

The workflow shape

In the expanded H3 Relay · LTX 2× Enhance graph, Rolling Input is the first thing built: it reads the current H3 segment and previous checkpoint, produces padded_images, then the graph runs VAE encode → 2x latent upscale → Rolling Inject → windowing/sampling → Rolling Checkpoint → tiled decode → Rolling Crop → Accept LTX. The rolling_context token is the accounting ledger that keeps the crop and accept steps honest about how much footage was real versus context versus padding.

You'll never place this node by hand - and don't. Its inputs are file paths pulled from the run's disk state, and the reference graph wires them for you. If you're debugging an enhancement that looks short or over-padded, check the status output of this node first; it tells you exactly how many frames went into the window and how many were context.

As with all the internal nodes: leave the wiring as built, keep run_name/shot_index consistent across the rolling family, and don't expect this to be useful outside a multi-shot LTX finish. Install the pack via ComfyUI Manager (H3 Relay) or git clone https://github.com/akatz-ai/h3-relay.git into custom_nodes; ComfyUI 0.32.0+, FFmpeg, and the LTX 2.5 files from MODELS.md are the requirements.