Text Encode with LTXV 2 System Prompt (Scaled Bias) (Legacy)

LTX 2 is the video model that made open-weights motion fast enough to feel real-time, and like the stills models it runs on an LLM text encoder - a Gemma-flavored one with its own <start_of_turn> template. This node is the LTXV 2 encoder: system prompt, scaled-bias emphasis, and - uniquely in this pack's encoder family - an optional image reference that goes through two separate paths, a VLM semantic path and a VAE structural path. That's the full modern video-conditioning stack in one node.

How it works

Core inputs: clip, prompt, system_prompt (optional). The node builds LTXV 2's Gemma template (<start_of_turn>system…<end_of_turn> plus the user turn). The scaled-bias layer means <tag=strength> in your prompt scales those token embeddings.

Then the reference image, which is what makes this more than a plain encoder:

• image (optional) - passed to the VLM semantic path as a soft image token inside the user turn.
• vae (optional) + vae_resolution - the same image, VAE-encoded as a reference latent (the structural path).
• ref_latent_mode - off (default) skips the latent entirely. single/multi append the latent(s) to the conditioning. parallel-single/parallel-multi run them in a separate conditioning stream so the reference structure doesn't semantically override your prompt - the distinction matters when you want composition guidance without the model copying the reference's content.
• vae_dimension_multiple (default 8) - pixel alignment for the reference before VAE encoding.

The tooltip says it plainly: MiniMax H3 users should keep this off and use Core's dedicated H3 reference conditioning instead - this node's reference path is LTXV-oriented.

Why two paths

The semantic path (image as soft tokens in the prompt) tells the VLM what the reference is. The structural path (VAE latent) tells the diffusion model where it is - composition, layout, geometry. Running both is how you get an image-to-video edit that holds the reference's structure, not just its subject. If you only wire image, you get semantics without structure; the parallel-* modes exist for when the structure would otherwise overpower the prompt.

Where people get burned

• No image = reference off, fine. The image input is optional; a prompt-only encode works normally.
• The parallel-* modes are for reference-heavy edits; on simple scenes the plain single mode can override your prompt's intent with the reference's content.
• LTXV 2's template - don't cross-feed a different model's clip.
• Bias strengths above ~1.5 oversaturate, same as the rest of the family.

Installing it

Ships in silveroxides/ComfyUI-UtilsCollection. ComfyUI Manager: search ComfyUI-UtilsCollection, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection
cd ComfyUI-UtilsCollection
pip install -r requirements.txt   # opencv-python, typing-extensions

Restart. No node-specific model downloads; LTXV 2 weights load normally. Legacy alias of UC_ScaledBiasTextEncodeLtxv2SystemPrompt, identical behavior.