LTX Unified Sampler (CRT)

LTX 2.3 is the Lightricks release that finally made the model competitive: 22B parameters, synchronized audio and video in one model, a rebuilt VAE with sharper detail, and the I2V that actually moves. It's genuinely the best local video model for a lot of people now. But "one model" doesn't mean one node - a proper LTX 2.3 workflow is a small forest of encoders, VAEs, guides, upscalers and samplers. This node's job is to fold that forest into a single box.

You hand it a models pipe and a config pipe from the pack's sibling nodes, pick a mode, and it runs the whole generation: conditioning, sampling, latent upscaling, refinement, VAE decode, and it hands you back images and the audio track.

How it works

The Unified Sampler implements LTX 2.3's two-stage default: generate at half resolution, run the latent upscaler, then a refinement pass. The hq toggle skips all that and does one full-resolution pass instead - maximum quality, longer run. Along the way it manages the pieces that make LTX 2.3 memory-friendly: low_vram unloads the text encoder after conditioning and the VAEs before sampling (they reload for decode), unload_model_before_vae_decode frees the diffusion model before decode, and vae_decode_tiled cuts peak VRAM during decode at the cost of time.

The mode switch drives everything:

• I2V - image-to-video from a first frame.
• T2V - pure text-to-video.
• V2V - video-to-video, with its own sub-modes: Depth Control (a depth guide conditions the output), Outpaint, or Upscale, tuned by v2v_guide_strength.

Audio is first-class here. frame_count_from_audio derives the video length from your source audio instead of the frame_count widget, and generated_audio_gain_db trims the synthesized audio track's level - negative defaults exist because LTX's generated audio tends to come in hot.

The inputs that matter

The two pipe inputs are non-negotiable: models_pipe (LTX23_US_MODELS_PIPE, from the pack's LTX 2.3 US Models Pipe node) and config_pipe (LTX23_US_CONFIG_PIPE, from LTX 2.3 US Config). Everything else is knobs:

• workflow_mode - I2V / T2V / V2V.
• hq - one-pass full-res vs. two-stage.
• megapixels_target and aspect_ratio - the output canvas, with a big list of presets.
• frame_count - length, default 161 (the LTX-friendly 8n+1 shape).
• steps - defaults to 9, right for the distilled model.
• depth_mouth_mask, mouth_mask_expand, mouth_mask_blur - lip-motion influence on the depth guide, for lipsync work. The three older mouth inputs (mouth_detect_megapixels, mouth_single_item, mouth_detect_chunk_size) are deprecated and ignored, per their tooltips - don't waste time tuning them.

Outputs are images (IMAGE) and audio (AUDIO). Wire them to a Save Video With Path-style saver.

Installing it

Install the pack: ComfyUI Manager → search CRT-Nodes, or git clone https://github.com/PGCRT/CRT-Nodes.git into custom_nodes plus pip install -r requirements.txt, then restart. The models themselves come from the AutoDL loader nodes (LTX 2.3 Model, VAEs, Gemma 3 + projection text encoders), which auto-download on first run.

Where people get burned

VRAM is the whole game. LTX 2.3 is 22B plus a 12B-class Gemma 3 text encoder - the KB's LTX history is a graveyard of OOM threads. Leave low_vram on, and if you still hit walls, the pack's GGUF Q4 model node exists precisely to get this running on smaller cards. Also: the deprecated mouth inputs are a classic trap because they look functional. They're compatibility shims, hardcoded internally - set depth_mouth_mask and the expand/blur values and ignore the rest.