Description
The LTX2 LoRA Loader Advanced node provides per-block strength control for applying LoRAs to LTX2 video diffusion models. Unlike a standard LoRA loader, this node allows you to independently adjust the influence of the LoRA on different types of attention layers within the model: video, video‑to‑audio cross-attention, audio, audio‑to‑video cross-attention, and all other layers not caught by those filters. This gives you fine‑grained control over how the LoRA modifies the model’s behaviour across modalities.
Inputs
| Input | Type | Required | Default | Min | Max | Step | Tooltip / Description |
|---|---|---|---|---|---|---|---|
lora_name | COMBO | Yes | — | — | — | — | The name of the LoRA (select from available LoRAs or type a name). |
model | MODEL | Yes | — | — | — | — | The diffusion model the LoRA will be applied to. |
strength_model | FLOAT | Yes | 1.0 | -100.0 | 100.0 | 0.01 | How strongly to modify the diffusion model. This value can be negative (inverts the effect). |
opt_lora_path | STRING | No | (empty) | — | — | — | Absolute path of the LoRA file. If provided, overrides the selected lora_name lookup. |
video | FLOAT | Yes | 1.0 | 0.0 | 1.0 | 0.01 | Strength for video attention layers. |
video_to_audio | FLOAT | Yes | 1.0 | 0.0 | 1.0 | 0.01 | Strength for video‑to‑audio cross‑attention layers. |
audio | FLOAT | Yes | 1.0 | 0.0 | 1.0 | 0.01 | Strength for audio attention layers. |
audio_to_video | FLOAT | Yes | 1.0 | 0.0 | 1.0 | 0.01 | Strength for audio‑to‑video cross‑attention layers. |
other | FLOAT | Yes | 1.0 | 0.0 | 1.0 | 0.01 | Strength for layers not caught by the other layer filters (e.g., shared or residual components). |
Outputs
| Output | Type | Description |
|---|---|---|
model | MODEL | The diffusion model with the LoRA applied, modulated by the per‑block strengths. |
lora_name | STRING | The name of the loaded LoRA (useful for logging or passing to other nodes). |
lora_path | STRING | The absolute path of the loaded LoRA file (or an empty string if only a name was used). |
Usage Notes
- The
strength_modelparameter controls the overall influence of the LoRA on the model. The per‑block strength parameters (video,video_to_audio, etc.) multiply against this overall strength for their respective blocks. For example, ifstrength_model = 1.5andvideo = 0.5, the effective strength on video attention layers is1.5 × 0.5 = 0.75. - A single per‑block strength of
0completely disables the LoRA’s effect on that block type, while a value of1.0applies the full (scaled) LoRA effect. - The overall
strength_modelcan be negative, which inverts the direction of the LoRA’s modification. Per‑block strengths should be kept in[0.0, 1.0]; negative per‑block values are not supported by the interface. - The
opt_lora_pathinput allows you to load a LoRA from a specific file path, bypassing the named‑LoRA selection. If bothlora_nameandopt_lora_pathare provided, the path takes precedence. - This node is designed specifically for LTX2 models. Using it with other model architectures may produce unexpected results.
Comments
Sign in with GitHub to join the discussion.