LTX 2.3 Generate All-in-One

Most LTX 2.3 workflows are a row of native nodes: CLIP encode, LTXVConditioning, ModelSamplingLTXV, scheduler, guider, sampler, crop guides, VAE decode. That's fine when you want to fiddle with each stage, and it's tedious when all you want is a clip. LTX 2.3 Generate All-in-One from the comfui-helto-ltx pack fuses the entire chain into a single node: prompts in, decoded images and an audio object out. No latent plumbing, no VAE decode node sitting at the end.

What it actually runs

This isn't a wrapper around a KSampler - it reuses ComfyUI's native LTXV components in sequence, and the README names them: CLIP text encode, LTXVConditioning, ModelSamplingLTXV, LTXVScheduler (or manual sigma parsing), RandomNoise, CFGGuider, SamplerCustomAdvanced, LTXVCropGuides, then VAE decode. Image guides go through the same native path as the pack's other nodes (append to latent, update conditioning with keyframe_idxs, guide_attention_entries, noise_mask), and crucially it crops the guide frames itself before decode - so no surprise extra frames.

The inputs that matter

Required are model (a native LTXV or LTXV AV checkpoint), clip, and vae. After those, the ones a beginner actually sets:

• positive_prompt / negative_prompt - plain strings, encoded inside the node. LTX likes long, specific prompts; short ones reliably underperform.
• num_frames (default 97) - pixel frame count, and yes, it must be 8*n + 1. The scheduler and any internally-created latent honor it.
• steps (default 30), cfg (default 1), sampler_name (default euler_cfg_pp) - the sampling settings. LTX 2.3 distilled is an 8-step-at-CFG-1 model; the dev model wants more steps.
• max_shift (2.05) / base_shift (0.95) / stretch / terminal (0.1) - the LTXV model-sampling and scheduler shift knobs. The defaults are the normal LTX starting points; leave them alone until you know you need to move them.
• seed - for reproducibility, or for the seed-hunting that LTX 2.3's prompt-adherence quirks make a habit.
• sigma_mode - ltx_scheduler (default) or manual. In manual mode you paste a comma-separated sigma schedule into manual_sigmas, and the number of denoise steps is (values − 1); steps, stretch, and terminal are ignored, though the shift values still patch the model.
• audio_mode - passthrough (default) or native_av. Passthrough just trims/pads any connected audio to the video length (or outputs silence). native_av is the interesting one: it samples video and audio together, which needs an LTXV AV model plus an audio_vae from the native LTXV Audio VAE Loader.

The optional start_images (IMAGE batch) works like the other nodes: a multi-frame guide at frame 0. The guides_json input is hidden and holds the guide data managed by the node's built-in image-guide UI.

Outputs and what they wire into

images is a plain IMAGE tensor - preview it, save it, or feed it into an upscale or V2V pass. audio is an AUDIO object (the passthrough track, or the decoded generated audio in native_av). The two outputs are the whole story; there's no latent to crop or condition to chase.

Install

From the same light pack - ComfyUI Manager, search "comfui-helto-ltx", or:

cd ComfyUI/custom_nodes
git clone https://github.com/helto4real/comfui-helto-ltx.git

Restart ComfyUI. No extra Python dependencies; the pack runs on ComfyUI's native LTXV support. The heavy lifting is on the model side: LTX 2.3 is a 22B model, and on consumer GPUs you'll want quantized weights plus serious system RAM (64GB is the community's comfort zone). For native_av audio you additionally need an LTXV AV checkpoint and its audio VAE.

Gotchas

• native_av fails. It requires both an LTXV AV model and an audio_vae. If you just want an external soundtrack alongside generated frames, use passthrough.
• Manual sigmas error. manual_sigmas needs at least two numeric values - nine sigmas = eight denoise steps.
• The clip is shorter/longer than expected. Check num_frames is 8*n + 1. LTX doesn't round quietly.

Reach for this node when you want a fast single pass and don't need per-stage control. For two-stage upscale work, step back to the Manager + Apply pair - those are built for routing one guide list through two sampler stages, which this convenience node deliberately isn't.