Workflow Config LTX2.3

Workflow Config LTX2.3 is the sibling of Workflow Config WAN2.2 with the guts swapped for Lightricks' model. Same movie / scene / take system, same JSON configs in user/default/workflows/.dx_mgr/, same "pick a scene, everything loads itself" payoff. But LTX 2.3 is a different beast, and the node shows it: there's no two-pass split, no timestep shift - instead there's a checkpoint and a standalone transformer, a separate audio VAE, and a distillation LoRA slot right up front where it belongs.

What's different about LTX 2.3 here

LTX 2.3 is the 22B release that made the model actually competitive locally - synchronized video and audio out of one model, real I2V motion, a cleaner vocoder. That "audio out of one model" is why the node has audio_vae and video_vae as separate outputs: the video and audio latents decode through different VAEs, and if you want sound in your clip, both have to be wired. It's also why there's a dedicated distillation_lora slot - the distilled checkpoint is a LoRA you stack on the dev model, and it's by far the most common thing you'll want applied.

The config editor stores a Checkpoint (model + CLIP + VAE in one file) or standalone UNet/Transformer, Video VAE, Audio VAE, and CLIP / CLIP 2 paths. Fill whichever set you use - both sets of outputs stay available.

Inputs and outputs

Inputs are the three selectors: movie, scene, take. The outputs you'll wire into the sampler:

• transformer_stack and checkpoint_stack - ready-to-use MODELs with all enabled LoRAs already applied. Pick one depending on which loading path you set up. Unlike the WAN2.2 node there's no shift built in - that's a WAN2.2-only feature, so you're on your own for any sampling tweaks.
• checkpoint_model / checkpoint_vae / checkpoint_clip and transformer_only / video_vae / audio_vae / dual_clip - the raw loaded components.
• image (reference input) and audio (a reference audio path decodes to the AUDIO output for downstream use).
• width, height, steps, seed, cfg, total_frames, fps - LTX frame counts need to land on "divisible by 8 plus 1", which is worth remembering if your scene keeps erroring at encode time.
• master_prmt, pos_prompt, neg_prompt, is_relay_prompt - same Smart / Beats / Simple prompt system as the WAN2.2 node, with the master prompt prepended for Simple and Beats.
• distillation_lora plus lora_2 through lora_8 (8 slots total), filename, is_t2v, flag_1–flag_3, and custom_1 / custom_2.

Installing it

Same pack, same install:

cd ComfyUI/custom_nodes
git clone https://github.com/denyazzolin/comfyui-daz-tools

Restart ComfyUI. No extra pip dependencies. The standalone transformer field supports GGUF, but that path needs the separate ComfyUI-GGUF pack installed - without it, .gguf files won't appear and running a GGUF scene raises a clear error.

Gotchas

The shared ones from its WAN sibling apply: back up .dx_mgr/ (your scenes are plain JSON, easy to lose), check that model files named in the editor actually exist, and remember Save vs + Version - Save overwrites the current take, "+ Version" snapshots a new one so you can A/B prompts without destroying the old take. If the editor UI ever looks wrong after a ComfyUI update, the Nodes 2.0 frontend toggle is the usual suspect; the legacy canvas is where custom UIs like this still behave.