Clarification on IC-LoRA fine-tuning with distilled models

[wymanCV]

Hi LTX team,

Thanks for releasing the LTX-2 trainer and IC-LoRA examples. I have a question about the recommended setup for fine-tuning IC-LoRA for LTX-2.3.

The pipeline docs mention that ICLoraPipeline can only be used with a distilled model, and inference seems to use the fixed distilled sigma schedule. However, the trainer's v2v_ic_lora.yaml uses the default flow-matching timestep sampler:

flow_matching:
  timestep_sampling_mode: "shifted_logit_normal"
  timestep_sampling_params: {}

Could you clarify the intended recipe?

1. Should IC-LoRA fine-tuning use the distilled checkpoint as model.model_path?
2. If so, should training timesteps/sigmas be sampled from the distilled inference schedule, or is the default shifted_logit_normal sampler still recommended?
3. Is there any expected mismatch from training over continuous random timesteps while inference only uses the fixed distilled sigmas?

Any guidance on the recommended LTX-2.3 IC-LoRA fine-tuning setup would be appreciated. Thanks!

[matanby]

Hi Wuyang LI (@wymanCV)!

The recommended recipe for training any LoRA (including IC-LoRAs) is to train on top of the LTX 2.3 base model, not the distilled one. Then, at inference time, you can apply the trained LoRA on either the base model or the distilled model (using the shorter distilled inference schedule). Both options are valid.