toobusy LTX2.3 Compact AV Sampler

If you've ever stared at a stock LTX-2 workflow, you know the AV sampling block: RandomNoise, LTXVConcatAVLatent, CFGGuider, KSamplerSelect, ManualSigmas, SamplerCustomAdvanced, LTXVSeparateAVLatent, and LTXVCropGuides - eight nodes that do nothing but sample the video+audio latent and split it back apart. toobusy LTX2.3 Compact AV Sampler is that whole block in one node. You feed it a model, both latents, and conditioning, and it returns a cropped, separated result.

How it works

This is the pack's "fold" philosophy at its purest: it calls the exact same ComfyUI native nodes internally, in the same order. Noise from your seed, video and audio latents concatenated into one AV latent, CFG-guided sampling through your chosen sampler along a sigma schedule, then LTXVSeparateAVLatent splits it back, and LTXVCropGuides trims the conditioning and latent to the sampled length. If you've already built the block manually, you're not learning anything new - you're just deleting eight nodes from your graph.

The inputs that matter

• model, positive, negative, video_latent, audio_latent - the normal suspects. The latents come from toobusy LTX2.3 Empty AV Latents (or the native LTX empties); conditioning from toobusy LTX2.3 Prompt Guide (or CLIPTextEncode + LTXVConditioning).
• cfg - defaults to 1.0, and for LTX 2.3 distilled that's the right baseline. The distilled model is CFG-1; nudge toward 1.1–2.0 only if you're hitting blown-out lighting.
• sampler_name - 44 samplers, default euler. For LTX 2.3's audio+video path the community standard is DDIM-uniform with euler around 8–12 steps, but this node is a sampler wrapper, not a scheduler, so the actual step count comes from your sigmas.
• manual_sigmas - the string of sigma values (default 1.0, 0.99375, 0.9875, ...) that defines your step schedule. This is the "9 steps" knob in disguise.
• sigmas (optional) - connect a SIGMAS output from a scheduler node and it overrides manual_sigmas entirely.

Outputs and where they go

• positive / negative - the cropped conditioning, ready for a second pass or a refine sampler.
• video_latent / audio_latent - separated and cropped. Video latent goes to a VAE decode, audio latent to an audio VAE decode.

Install and gotchas

Pack install is the same as the rest of the toobusy nodes - ComfyUI Manager search "toobusy", or clone into custom_nodes and restart. The hard requirement is ComfyUI's LTX 2.3 node set (LTXV*). This node literally calls those classes by name, so on a build without LTX 2.3 support it fails at runtime with a clear "required node not available" message - read that error as "update ComfyUI," not "fix your workflow."

The subtle thing to remember: this node samples, it doesn't decide quality policy. LTX 2.3's audio can come out mumbled if you're using the wrong distilled LoRA revision (the community moved from 1.0 to 1.1 for exactly that reason), and that's upstream of this node. Get the conditioning and the latent right, and the Compact AV Sampler is the boring, reliable middle of your graph - which is the point.