LTX2_SM_ENCODER

This is the node that actually talks to the model. LTX2_SM_ENCODER takes the Gemma 3 text encoder you loaded with LTX2_SM_Clip, runs your prompt (and negative) through it, and produces the positive / negative conditioning that the KSampler consumes. If you've used Comfy's CLIP Text Encode, this is the same shape - but LTX is pickier about what you feed it than SD ever was.

What it does

encoder_text() passes both prompts to the Gemma encoder and concatenates each one's video encoding with its audio encoding into a single 6144-wide embedding. Notice the trick in the code: if the negative prompt is empty, negative comes back as None - because in distilled mode there is no negative prompt at all (CFG is 1, no classifier-free guidance). So the negative output is only meaningful when you're running a non-distilled/dev pipeline.

The inputs that matter:

• enhance_prompt (default off) - toggles Gemma's built-in prompt enhancer, which rewrites your prompt before encoding (optionally using an image, if you feed images). The community consensus on LTX-2/2.3 is blunt: turn this off and write the prompt yourself. The built-in enhancer produces basic prose, and a long, detailed, self-written prompt beats it every time. LTX wants paragraphs - describe the shot, the motion, the camera, even the audio ("she begins to sing in English: …").
• save_emb (default on) - writes the embeddings to disk. This is the pack's killer feature: with it on, you can bypass this whole node on later runs and the KSampler will quietly re-read the saved embeddings (it calls read_lat_emb() when no conditioning is wired in). Encode once, seed-hunt forever. Given Gemma 12B is the slowest, heaviest thing in the graph, this turns a per-run tax into a one-time cost.
• streaming_prefetch_count (default 1) - layer-streams the text encoder to keep VRAM down. Leave it at 1 on small cards; 0 disables streaming.
• prompt / negative_prompt - the author's default is a long singing-puppet scene; it's there as an example of how detailed LTX wants things, not as a template you should keep.

Outputs are positive and negative conditioning into the KSampler. There's also an optional images input that only does anything when enhance_prompt is on (it feeds the image to Gemma for enhancement).

Installing & troubleshooting

Same pack install as every node here - ComfyUI Manager (search "ComfyUI_LTX2_SM") or:

cd ComfyUI/custom_nodes
git clone https://github.com/smthemex/ComfyUI_LTX2_SM.git
cd ComfyUI_LTX2_SM && pip install -r requirements.txt

Real-world pain points:

• Encoding is slow and it's the text encoder's fault, not yours. Gemma 3 12B is huge. Run it on CPU via the infer_device on LTX2_SM_Clip, and keep save_emb on.
• No negative output when you want one? In distilled mode, negatives are structurally absent - that's the model, not a bug. If you need real negatives, switch the Model node to a non-distilled sampling mode.
• "It ignores half my prompt." LTX-2.3's prompt adherence is still its weakest axis; the fix is writing more specific motion and camera instructions and seed-hunting, not longer run-on sentences. And per the 2.3 playbook, voice lines placed later in the prompt get spoken later in the clip.