distilled LoRA 450

Question

distilled LoRA 450

Accepted Answer

The term **distilled LoRA 450** refers to an official distillation adapter designed for the **LTX-2.5** 22B video generation model.

### Key Details
*   **Purpose:** It is used to apply "distilled-style" fast inference to the LTX-2.5 **dev** model. By applying this LoRA, users can achieve the faster sampling speeds associated with distilled models while maintaining the higher quality ceiling of the original dev checkpoint.
*   **Rank:** The "450" indicates the rank of the LoRA adapter (450), which is an increase from the rank 384 LoRA used for the previous LTX-2.3 version.
*   **Usage:**
    *   **Placement:** It is typically placed in the `ComfyUI/models/loras/` directory.
    *   **Compatibility:** It is specifically designed to be paired with the LTX-2.5 dev model. It should not be applied to an already-distilled checkpoint, as this can result in "double-distillation" and degraded quality.
    *   **Inference Settings:** For optimal results, it is generally used with specific inference schedules (often 8-step sampling) and a classifier-free guidance (CFG) scale of 1.0, consistent with distilled inference profiles.

In practice, this LoRA allows for efficient, high-quality video generation by bridging the gap between the flexible, high-fidelity "dev" models and the optimized, low-step "distilled" models.