ID-LoRA Prompt Encoder

Every other node in this pack is plumbing. This is the one where you actually type. The ID-LoRA Prompt Encoder sits between the loader and the sampler in both pipelines - one-stage and two-stage - and turns your text into the conditioning tensors the model denoises against.

Inputs and output

Three inputs, one of which is already filled in for you:

• pipeline - from either Model Loader. This is the one node both pipelines share.
• prompt (multiline) - your positive prompt.
• negative_prompt (multiline) - pre-filled with a long, sensible list of quality negatives: blur, artifacts, mismatched lip sync, robotic voice, added dialogue, and so on. The author wrote your negative prompt for you, which is a nice touch. You can keep it as-is forever.

Output is a single ID_LORA_CONDITIONING object carrying four tensors - positive and negative, each split into video and audio encodings - wired straight into the sampler.

The prompt format is the real feature

ID-LoRA expects a structured, three-section prompt:

[VISUAL]: A medium shot of a young man with curly brown hair,
sitting on a beige couch, speaking warmly.
[SPEECH]: We are proud to introduce ID-LoRA.
[SOUNDS]: The speaker has a moderate volume and conversational tone.
Light instrumental background music plays softly.

All three sections are optional but recommended. [VISUAL] describes the scene and the face, [SPEECH] is the actual lines the speaker delivers, and [SOUNDS] covers tone of voice and ambience. LTX as a family is famously prompt-hungry - short prompts reliably underperform, and the community's standing advice is long, specific prose over tags - and this labeled format is the pack's answer to that. If your speaker sounds wrong, the [SOUNDS] section is the first place to look.

How it works

Encoding is not free, and the pack goes out of its way to make it cheap. The node loads the Gemma 3 12B text encoder - the ~23GB one, the same encoder that caused most LTX-2 launch-day OOM errors - encodes both prompts, then explicitly frees the encoder and garbage-collects before returning. That's deliberate sequencing: prompt encoding happens before the sampler loads the 22B transformer, so the two biggest models never occupy VRAM at the same time. It's the reason this pack could run on 24GB cards at all, and it's the design that lets the loaders defer their heavy load_models() call.

Installing and troubleshooting

Shared pack install - see the ID-LoRA Model Loader article. At runtime this node is lightweight; the pain is all upstream. The Gemma download is ~23GB and needs a HuggingFace login plus license acceptance, and if transformers isn't pinned to 4.x the encode step can crash on a 5.x incompatibility before you ever see a single frame. Worth repeating: keep this pack on a ComfyUI install you don't mind pinning, because the recommended route - native LTXVReferenceAudio in core - is where everything new is happening.