Workflow Config LTX2.5

LTX is the speed tier of local video: where Wan burns minutes per clip, LTX burns seconds, and the whole community workflow built around that is iteration - generate ten takes, keep one. But every iteration means swapping models, VAEs, prompts, LoRAs, resolutions, seeds. Individually each swap is thirty seconds of clicking; collectively it's why you dread changing anything. Workflow Config LTX2.5 stores named "scenes" - a full snapshot of the video pipeline - in a JSON file, and reloads everything the moment you pick one from a dropdown.

It's built for the LTX 2.5 generation of the model, and the config reflects what that family needs. LTX-2 and later made synchronized audio and video their whole bet, so there are two VAEs: video_vae and audio_vae. The text encoder is the single Gemma-based file LTX 2.5 ships (gemma-12b-with-proj-…), which is why this node - unlike the pack's LTX2.3 variant - has one CLIP slot and no checkpoint loader. One CFG, 8 LoRA slots, and an optional latent upscaler round it out. The first LoRA slot is even named distillation_lora: LTX's distilled checkpoints run at 8 steps, CFG 1, so that slot is your speed knob by design.

The mechanics are the pack's shared scene system. Scenes live as dx_*.json "movie" files in ComfyUI/user/default/workflows/.dx_mgr/. Each movie holds multiple scenes; each scene holds numbered takes - versions of the same setup with different prompts, LoRAs, reference image, whatever. At runtime you only touch three inputs: movie, scene, take, all dropdowns. Everything else comes out of the outputs, so you wire the graph once and switching takes is just changing dropdowns.

Outputs worth knowing:

• transformer_stack (MODEL) - the transformer with all enabled LoRAs applied. This is what you wire into the sampler; it's the whole point.
• transformer_only (MODEL) - the bare model, if you'd rather load LoRAs yourself.
• video_vae / audio_vae (VAE), clip (CLIP) - the decode and text-encode stack.
• image (IMAGE) and audio (AUDIO) - reference inputs for I2V and sound conditioning.
• width, height, steps, seed, cfg, total_frames, fps - straight into the sampler and latent setup.
• pos_prompt, neg_prompt, master_prmt, is_relay_prompt - the prompts, with the relay flag on when a Smart (segmented) prompt is active.
• latent_upscaler (LATENT_UPSCALE_MODEL) - optional, loaded from models/latent_upscale_models; empty if you leave it on none.
• flag_1/2/3, custom_1/2, filename - routing and output-path helpers.

Editing happens in a full-screen panel (double-click the node, or hit Edit Take): a visual prompt segment editor, a sizing dialog that checks divisibility by 32, and duration quick-buttons that compute frames as duration × fps + 1. Save or "+ Take" writes back to disk.

Installation:

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

Restart, or use ComfyUI Manager and search comfyui-daz-tools. No pip requirements. If you plan to run GGUF-quantized transformers - the dropdown lists .gguf files alongside regular ones and loads them through ComfyUI-GGUF automatically - you need ComfyUI-GGUF installed, or GGUF entries won't even show up.

Real-world notes. The LTX-2 family's Gemma text encoder is huge - roughly 23GB of system RAM on top of the model - and the launch era was full of OOMs on 24GB cards without --reserve-vram. Budget memory before you blame the node. Your scenes are only as safe as the .dx_mgr folder; back it up or you're rebuilding everything. And if a take has Randomize seed on, the node rewrites the seed into the file on every run, which is why it constantly shows as "changed" - expected, not a bug. Finally, Smart prompts are relay payloads here; if you don't run a relay node downstream, keep prompts on Simple.