Question: custom LoRA trained on LTX 2.3 dev BF16 for distilled 1.1 FP8 inference
Hi Lightricks team,
I am preparing a custom character/identity LoRA for LTX 2.3 and would like to confirm the supported training/inference path before paying for cloud training.
My local ComfyUI inference workflow uses:
- LTX 2.3
- 22B
- distilled 1.1
- transformer-only
- FP8 scaled
- safetensors
The exact local inference transformer is:
ltx-2.3-22b-distilled-1.1_transformer_only_fp8_scaled.safetensors
Based on the official ltx-trainer documentation, the intended training base appears to be:
ltx-2.3-22b-dev.safetensors BF16
I also understand that the trainer requires:
google/gemma-3-12b-it-qat-q4_0-unquantized
My dataset is a character/identity dataset made of still images using F=1 samples, with clean captions. My plan is to provide the trigger token through the trainer/preprocessing option rather than manually inserting it into every caption.
My main concern is not only whether LoRA keys can technically load, but whether a custom identity LoRA trained on the BF16 dev checkpoint is expected to behave correctly when applied to the FP8 scaled distilled-1.1 transformer-only checkpoint.
Questions:
Is the path “dev BF16 training → custom LoRA → distilled 1.1 FP8 scaled inference” officially supported or expected to work?
For a custom LoRA intended for distilled 1.1 FP8 inference, should the training base be dev BF16, a distilled BF16 checkpoint, or another checkpoint?
Is there an official recommended recipe for character/identity LoRA training on LTX 2.3 22B, especially for still-image F=1 datasets?
Can a LoRA trained on the full BF16 dev checkpoint be applied to the transformer-only FP8 distilled-1.1 checkpoint without expected quality or behavior issues?
For ComfyUI inference, is a standard LoRA loader expected to work with this kind of LTX LoRA, or should a specific ComfyUI-LTXVideo LoRA loader/workflow be used?
What is the safest checkpoint choice before committing to cloud training?
Thank you for any official guidance, recommended config, or known limitation you can share.
Hi,
The recommended path is to train the LoRA on the LTX-2.3 22B Dev BF16 checkpoint, then apply it during inference on the Distilled 1.1 checkpoint.
This is the intended workflow for custom LoRAs targeting distilled inference. Training directly on the distilled checkpoint is not currently the recommended approach.
For character/identity LoRAs, we do not have one universal recipe. Results depend heavily on the dataset, captions, image diversity, and the identity being trained. F=1 still-image samples can be used, and supplying the trigger token through the trainer/preprocessing configuration is fine. The default hyperparameters in the reference trainer configs are a good starting point; you should expect to tune them based on validation results.
A LoRA trained against the full BF16 Dev model is expected to be usable with the transformer-only FP8-scaled Distilled 1.1 checkpoint, assuming the checkpoints and model configuration are from the same LTX-2.3 family. FP8 quantization and transformer-only packaging are inference optimizations and do not change the recommended training base. Nevertheless, identity fidelity and optimal LoRA strength should be validated on your specific dataset.
For ComfyUI, use the LTX-specific ComfyUI-LTXVideo nodes/workflows and LoRA loader. Instructions are available here:
https://docs.ltx.io/open-source-model/usage-guides/lo-ra#using-loras