⭐ Star LTXV All-in-One (2-Pass)

The workflow that most people use for LTX 2.3 - the one the community calls the Sulphur workflow - is a tall stack: load the transformer, load two text encoders, load two VAEs, load a latent upscaler, run pass one at half resolution, upscale the latent, run pass two at full resolution, decode. Star LTXV All-in-One (2-Pass) collapses that entire tower into a single node. It's the pack's flagship LTX node, and it's the reason a lot of people install StarNodes in the first place.

It's in ⭐StarNodes/Video, labeled for LTX 2.3 (the node's model dropdown expects the LTXV 2.3 A/V checkpoint family - sulphur2Base and friends).

How it works - the two-pass trick

LTX 2.3's quality comes from a two-stage render, and this node is a faithful port of it:

1. Pass 1 at half resolution - a fast coarse pass that establishes motion and composition on a small latent (cheap).
2. 2× latent upscale - a spatial upscaler model (ltx-2.3-spatial-upscaler-x2) doubles the latent between passes, not a pixel upscaler.
3. Pass 2 at full resolution - a refinement pass over the upscaled latent that adds the detail the final output needs.

Both passes use their own sampler (sampler_pass1 defaults to euler_ancestral_cfg_pp, sampler_pass2 to euler_cfg_pp), share one seed, and one cfg (default 1.0 - correct for the distilled 2.3 checkpoints). The sigma schedules come from the original workflow's note node: sigma_preset offers the three baked schedules (8 / 12 / 16 steps) or custom, where you paste your own into custom_sigmas_pass1. The second pass uses its own sigmas_pass2 (default 0.85, 0.725, 0.6, 0.4219, 0.0).

Modes and inputs

Three modes on the mode dropdown:

• text_to_video - prompt only.
• image_to_video - connect an image as the start frame.
• image_audio_to_video - connect image and an audio track, which is trimmed to the video length and preserved.

The node loads everything internally - base model, two text encoders (clip_1 for the main Gemma-3 encoder, clip_2 for the LTX text projection), video VAE, audio VAE, and the latent upscaler - and caches them so a big file is only re-loaded when the dropdown selection actually changes. There's a 3-slot LoRA stack (lora_1 … lora_3, the distilled LoRA convention of ~0.6 on slot one), and an optional model_override if you want to feed in an externally-patched model instead of the dropdown load.

video_size (HD ~1280px / FHD ~1920px / Custom) with ratio presets, and ratio_from_image (default on) snaps the aspect to your connected input image. seconds snap to the LTX-friendly 8n+1 frame pattern.

Outputs

images (the frame batch), audio (the generated soundtrack - this node can generate audio alongside video, which is LTX 2.3's party trick), and frame_rate as a float for downstream video nodes.

Installing it

StarNodes installs via ComfyUI Manager (search Starnodes, Install, restart) or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

Search the canvas for star after restart.

The reality check

This is a genuinely impressive consolidation, but know what you're signing up for. LTX 2.3 is a 22B model - even distilled and quantized it's a heavy load, and the two text encoders plus two VAEs plus an upscaler make this node memory-hungry by construction. The author ships sensible defaults so it works, but on a 12GB card expect offloading and patience, not magic. And because so much is hidden inside one node, when it fails it can be hard to see why - the console is your friend. Check that your model files actually landed in models/diffusion_models, models/text_encoders, models/vae and models/latent_upscale_models, since the dropdowns populate from those exact folders. If a dropdown is empty, that's a file-folder problem, not a node problem.