LoRA trained on base model vs distilled model

[sarlea1]

Hi, I have a few questions about the relationship between the base model and the distilled model, specifically around
LoRA training and inference compatibility.

Background

From reading the code, I understand there are two separate architectures:
* DistilledPipeline: uses ltx-2.3-22b-distilled.safetensors directly as a standalone checkpoint, running 8-step
inference via DISTILLED_SIGMA_VALUES
* TI2VidTwoStagesPipeline: uses the base model for Stage 1 (full CFG, ~30 steps), then applies a separate
distilled_lora on top of the base model for Stage 2 super-resolution refinement

Questions
1. Is ltx-2.3-22b-distilled.safetensors fine-tuned from the base model, or independently trained? This affects
whether a LoRA trained on the base model can be transferred to the distilled model.
2. If I train a LoRA on the base model and load it into DistilledPipeline, is the LoRA expected to work? Given that
DistilledPipeline uses 8-step distilled sigma values ([1.0, 0.99375, ..., 0.0]), but the LoRA was trained with the
standard full-range flow matching objective, I'm concerned about the mismatch.
3. What is the recommended workflow for training a custom LoRA that can be used with DistilledPipeline for fast
8-step inference? Should I:
* Train on the distilled model directly (using the standard flow matching training objective, not the distilled
sigma schedule), or
* Train on the base model and hope the weights are close enough for transfer?
4. The trainer does not seem to have any distillation-aware training mode. Is there a plan to support training
LoRAs specifically aligned with the distilled sigma schedule?

Any guidance on the intended workflow here would be greatly appreciated. Thanks!