Baberg/ltx-2.5-ic-lora-cartoon-dataset

LTX-2.5 IC-LoRA cartoon dataset — 27 paired clips

Training data for Baberg/ltx-2.5-22b-ic-lora-cel-character.

A fully synthetic dataset, created from scratch — nothing filmed or scraped. Cinematic live-action clips were generated, each built around one clearly identifiable central character. For each clip the first frame was extracted, the central character in it was replaced with a 2D cel-animated version of the same character in the same pose and framing, and the clip was then re-animated from that edited frame. Both halves of every pair therefore come from the same shot — identical framing, camera motion and timing — and differ only in how the character is rendered.

27 pairs are included here; the released LoRA was trained on 20 of them. See EXCLUDED.md for which seven were dropped and why — that selection step was the single change that fixed the model's worst failure mode.

Captions were generated with Google Gemini 3.5 Flash (gemini-3.5-flash) on the reference halves; see caption_instruction.txt for the exact instruction used.

Each pair is the same shot twice: a live-action reference, and a target in which one central character has been converted to hand-drawn 2D cel animation while everything else stays photographic. The character's reflections are also stylised, and the targets add a drawn cast shadow on the ground.

Layout

videos/training_20/  40 conformed clips — the 20 pairs the LoRA was trained on
videos/excluded_7/   14 conformed clips — the 7 pairs dropped (see EXCLUDED.md)
                     all 768x448, 24 fps, 97 frames, silent
                     <stem>_ref.mp4    = live-action (IC-LoRA conditioning input)
                     <stem>_target.mp4 = 2D-converted (generation target)
raw/reference/       27 original live-action clips (~720p, 121 frames)
raw/target/          27 original converted clips
dataset.json         trainer metadata for all 27 pairs
dataset_training_20.json  trainer metadata for the 20 training pairs only
captions.json        raw captioner output
conform_report.json  per-pair conforming record incl. residual FOV delta
caption_instruction.txt  the exact captioner instruction used
scripts/             conform, verify and alignment-audit scripts

dataset.json uses the LTX trainer's column convention:

{"video": "videos/<stem>_target.mp4", "reference_video": "videos/<stem>_ref.mp4", "caption": "..."}

Conforming

Raw clips are ~720p / 121 frames; the trainer bucket is 768×448×97 (97 = 8n+1). Two defects in the raw data had to be handled, and both are worth knowing if you re-derive the conformed set:

1. 12 targets declare r_frame_rate=60/1 but are really ~24.12 fps VFR (avg_frame_rate 7260/301) carrying the same 121 frames as their reference. Honouring those timestamps makes ffmpeg resample 5:2 and silently destroys frame-for-frame pairing. The fix is -r 24 before -i, which reinterprets the input as CFR 24 and discards source timestamps. scripts/audit_alignment.py verifies the result: all 54 clips map 1:1 to their source (lag 0, RMSE ≈ 0.6 = encoding noise).
2. Aspect ratios differ within pairs (1.7645–1.7961). Both halves are cropped to full height and a centred 768:448 window, then scaled — the largest field of view the two halves can share without stretching. Residual mismatch 0.83 % worst case, 0.00 % for 7 pairs; per-pair figures in conform_report.json.

Reproduce with:

python scripts/conform_pairs.py --raw-root raw --out-dir videos
python scripts/verify_dataset.py --videos-dir videos     # exits non-zero on any mismatch
python scripts/audit_alignment.py                        # proves 1:1 frame mapping

Captions

Generated with the LTX trainer's own caption_videos.py -c gemini_flash (gemini-3.5-flash), run against the reference (live-action) half of each pair. The instruction is the trainer's DEFAULT_VIDEO_CAPTION_INSTRUCTION with the audio bullets removed (this is a video-only LoRA on silent clips) and a requirement added that the caption state the footage is live-action and name one central character distinctly. Exact text in caption_instruction.txt.

All 27 captions contain the phrase "live-action footage"; 62–120 words (mean 90); 17 of 27 also name additional people, so the captions carry disambiguating information about which character is the subject.

Notes

• No held-out split. All 27 pairs were used for training.
• The converted character is not always the most prominent person — in 24_poker-single-lamp it is the elderly woman further back, not the man closest to camera.