Personal high-quality pipeline (two-stage LTX-2.3 diffusion + distilled refinement), much slower and more expensive per clip than Wan2.2. Targeting 1152x2048 for clips up to ~4.8s (41/81/121 frames) as of the latest change - not yet confirmed to fit on the A100-80GB this runs on at that size, so the first few generations at this length are effectively a live test; if it OOMs, try the shortest length (41 frames) first. Longer single segments (161/201 frames) still step down to 896x512 / 832x448, unchanged for now. The "Generated with" caption under the result always shows the exact resolution that actually ran, so you can confirm it directly instead of guessing.

Segments chain together into one continuous clip, so Length is how many. Each new segment only sees the last frame of the one before it, not the full motion — past ~5 segments (~17s), drift in color/detail/character consistency gets more noticeable toward the end. Leave Frames per segment at 81 unless you know you want shorter/longer individual segments.

Each segment here is a full ~5–10 min generation on its own — chaining N segments multiplies both generation time and GPU cost by N, so keep this conservative. Segments chain via last-frame continuity, same as Wan2.2 (an attached extra reference image, see the + button, overrides a later segment's starting frame instead).