LTX Chunk Assembler (TJ 2.5)

Joins the chunks of a run into one continuous PNG sequence, cross-dissolving the overlapping frames. Writes to <run>/final/ rather than returning an IMAGE batch: a 16s 1920x1088 sequence is about 10GB as float32, which is not something to hold in memory just to hand to a video encoder.

By designloves2·Created 4 months ago·Updated 3 days ago· 35

final_folder
info

◄run_namerun01►

◄runs_folderltx25_one_runs►

◄overlap_frames8►

◄crossfadetrue►

◄write_videotrue►

◄video_fps25.00►

◄video_crf18►

◄match_levelsfalse►

◄audio_crossfadeauto►

◄write_final_framesfalse►

◄frame_formatjpeg►

◄jpeg_quality95►

◄seam_modeearly_cut►

◄seam_blend_frames3►

CategoryWhatDreamsCost CS 2.5

Inputs (14)

Name	Type	Default	Description
run_name	STRING	run01	The run to assemble. Must match what the Chunk Writer used.
runs_folder	STRING	ltx25_one_runs	Subfolder of ComfyUI's output directory holding the run.
overlap_frames	INT	80–64	How many frames each chunk shares with the next. Must match the handoff_frames you rendered with, and you must have started each window that many frames before the previous chunk ended.
crossfade	BOOLEAN	true	Blend across the overlap instead of cutting. Off = hard cut, overlap frames taken from the earlier chunk.
write_video	BOOLEAN	true	Also encode <run>_final.mp4 alongside the final/ PNG sequence.
video_fps	FLOAT	25.001–240	Frame rate for the assembled video. Match your timeline.
video_crf	INT	180–51	h264 quality. Lower is better and bigger; 18 is visually near-lossless.
match_levels	BOOLEAN	false	Offset each chunk's brightness to match the one before it. This averages the MEAN over every overlap frame and never touches contrast - unlike the writer's per-frame match, which was unreliable. Still off by default: try the crossfade alone first.
audio_crossfade	COMBO	auto	How to blend audio across the overlap. Auto measures each seam and picks: identical material (an inpainted song) passes through untouched with a linear fade; independent generated sound gets equal_power so the level holds; music the model rendered slightly differently in each chunk gets a short 15ms splice, because a long blend of near-but-not-identical audio flanges. Short forces that splice everywhere, off is a hard cut. Every setting consumes the same overlap, so the audio never drifts out of sync with the picture.
write_final_frames	BOOLEAN	false	Also write the assembled sequence as stills in <run>/final/. Off by default because the mp4 is the deliverable. Forced on if write_video is off.
frame_format	COMBO	jpeg	Format for the final/ stills, when they are written at all. Chunk frames are read in whichever format the writer produced.
jpeg_quality	INT	9570–100	Only used when frame_format is jpeg. Chroma is never subsampled.
seam_mode	COMBO	early_cut	Where to join two chunks. The next chunk's FIRST frame is the handoff guide reproduced, so it is pixel-identical to the previous chunk's frame at (end - handoff_frames). early_cut switches exactly there - identical content, nothing mixed, no ghost. early_scurve blends over seam_blend_frames with an S-curve - soft at both ends, so a difference in camera velocity between the chunks is ramped rather than stepped. early_blend is the same length with a straight fade. dissolve and hard_cut are the old behaviour, which joined at the END of the overlap after the chunks had already diverged. This setting supersedes the crossfade toggle above.
seam_blend_frames	INT	31–32	How many frames the early_scurve / early_blend seam covers. Clamped to handoff_frames. Short is usually right: the chunks only agree near the start of the overlap, so a long blend reaches into frames that have already diverged.

Outputs (2)

Name	Type	Description
final_folder	STRING	—
info	STRING	—