Parses timeline JSON, loads/trims audio directly from memory using PyAV, and aligns to a global timeline yielding ComfyUI's format. Output le

Apply H.264 compression artefacts to an [N, H, W, 3] float32 tensor (ComfyUI image format). crf=0 means no compression. Uses PyAV to encode/decode

Convert pixel-space segment lengths to integer latent-space lengths using the largest-remainder method. Targets the full `latent_frames` when the

Decode an image from the ComfyUI input folder (if imageFile provided) or fallback to base64 to a ComfyUI-style image tensor of shape [1, H, W, 3],

Extracts a sequence of frames from a video file based on the segment's trim parameters, and returns them as an [N, H, W, 3] float32 tensor.

Wrap an existing LTX cross-attn forward (the default `CrossAttention.forward` or another node's patch — e.g. KJNodes NAG), injecting PromptRelay's

transformer_options override that routes mask-bearing attention calls through attention_pytorch (sage/etc. drop arbitrary masks). Chains to a prio

Per-segment metadata for the temporal penalty. relay_options (optional dict) overrides per-stream knobs: video_strength, video_window_sca

Gaussian penalty matrix [Lq, Lk] for video cross-attention (integer frame indexing).

Penalty matrix for queries that don't map to integer frames (e.g. LTXAV audio tokens).

Closure: mask_fn(Lq, Lk, dtype, device, transformer_options) -> additive mask or None. Takes shapes/dtype/device instead of tensors so callers ca

Return (arch, patch_size, temporal_stride) for latent geometry. temporal_stride is the VAE's pixel→latent temporal compression factor, used t

Validate or auto-distribute segment frame counts, capped to fit within latent_frames.

Convert audio to float 32 bits PCM format.

Extract the raw SPiece/HF tokenizer from a ComfyUI CLIP object.

Uses the latest ComfyUI av-based decoding for maximum compatibility.

Tokenize global + space-prefixed locals; return (full_prompt, per-local token ranges). Uses incremental tokenization to avoid SentencePiece conte