feat: add LTX-2.5 flexible multi-modality training

[NancyFyong]

What

Adds LTX-2.5 multi-modality flexible training to DiffSynth-Studio, following the repository's existing LTX-2/2.3 training flow. Built on top of the LTX-2.5 split-component inference pipeline.

Supports the public flexible conditioning modes:

• T2AV / I2AV (first-frame)
• Video prefix/suffix extension
• A2V / V2A (Foley)
• Video inpainting / outpainting
• T2A
• Audio prefix/suffix extension, audio inpainting
• V2V / A2A / AV2AV IC-LoRA

Each run can use the Dev or Distilled split BF16 DiT, as LoRA or full-DiT fine-tuning (full fine-tuning via an 8-process BF16 DeepSpeed ZeRO-3 accelerate config).

Implementation

• diffsynth/pipelines/ltx25_training.py: training-only LTX25FlexibleTrainingPipeline and config. Per-modality is_generated, VAE encoding of target/reference media, intrinsic conditioning masks, reference latent positions, and modality-selective masked flow-matching loss.
• LTXModel.forward: supports an absent video modality so audio-only (T2A / audio extension) training works, preserving existing dual-modality inference.
• LTX25AudioVideoPipeline.from_pretrained: subclass-safe classmethod so the training pipeline can reuse split-component loading.
• examples/ltx2/model_training/train_ltx25.py: generic launcher with cache (sft:data_process) and DiT-only (sft:train) split, LoRA and full-DiT options.
• examples/ltx2/model_training/ltx25/*.yaml: 15 per-mode training-strategy configs.
• examples/ltx2/model_training/{lora,full}/LTX-2.5-flexible-splited.sh + full/accelerate_config_ltx2_5_deepspeed_zero3.yaml: LoRA and full-DiT recipes.

Validation (no download / no real training on large data)

• All new Python sources compile; launcher --help works.
• All 15 mode configs parse and pass config validation.
• Dev/Distilled split ModelConfig composition builds for cache and train tasks.
• CPU dummy DiT validates video-only, audio-only, frozen-cross-modality, reference-token, masked-loss, and backward paths.
• LoRA and full (ZeRO-3) T2AV runs completed on the bundled example dataset (121-frame video + audio) over 10 epochs, with TensorBoard loss logging and epoch/step checkpoints saved.
• LTX-2.5 DiT 4,091-key metadata contract and inference behavior unchanged.

Use

LoRA:

bash examples/ltx2/model_training/lora/LTX-2.5-flexible-splited.sh t2av dev

Full DiT (DeepSpeed ZeRO-3):

bash examples/ltx2/model_training/full/LTX-2.5-flexible-splited.sh t2av dev

LTX25_MODEL_ROOT, LTX25_DATA_ROOT, LTX25_METADATA, LTX25_NUM_EPOCHS and LTX25_DATASET_REPEAT are configurable.

Note

This draft stacks on the LTX-2.5 inference PR (upstream modelscope#1602, fork ltx2_5). Once that lands on main, this PR's base can be switched to main to show only the training diff.