LTX 2.3 / createVideo

This node runs LTX 2.3 - Lightricks' 22B model - on Civitai's cloud, from inside ComfyUI. It's the flagship of the LTX 2.3 family in civitai/civitai-comfy-nodes (the others are audioToVideo, editVideo, extendVideo, firstLastFrameToVideo, and videoToVideo), and it does text-to-video plus optional image-to-video through the images input. No weights to download, no 22GB Gemma-class text encoder to offload, no VRAM math. Just Buzz.

The local LTX 2.3 story is genuinely good - the rebuilt VAE fixed the blurry 2.x output, I2V stopped freezing, native portrait mode up to 1080x1920 shipped, and by mid-2026 a real share of the community called it the best video model you can run locally (the KB's ltx-video doc has the full arc). What that story skips is the hardware: official minimum is 32GB+ VRAM, and practically everyone runs quants with 64GB of system RAM behind them. This node hands you that exact model with none of the setup, at the cost of per-run billing and your inputs leaving the machine.

How it works

Same orchestration plumbing as the whole pack: assemble a videoGen workflow with the LTX 2.3 engine, submit to the Civitai Orchestration API, long-poll until the job succeeds, download the MP4 as a native VIDEO. The interesting bit is that Civitai runs a Comfy workflow server-side - the tooltips reference ComfyUI graph details like ComfyLtx23VideoGenInput.Model and an upscale-LoRA stage - so you're renting Civitai's ComfyUI plus LTX 2.3, with the results shipped back to your graph.

Inputs and outputs that matter

• prompt (required) - the clip. LTX 2.3 is still weakest on prompt adherence even after the 2.3 improvements, so write camera, motion, and subject explicitly; the local community's "long prompts and seed hunting" advice carries over.
• model - 22b-dev (default) or 22b-distilled. Dev follows prompts better and needs more steps; distilled is faster and more stable but dumber. Locally the distilled LoRA 1.1 is the version to use - Civitai's engine defaults are a reasonable proxy, but you can't stack the LoRA yourself here.
• width / height - default 1280×720. LTX 2.3 requires width/height divisible by 32; the hosted engine should handle rounding, but feeding off-grid values is asking for a server-side rejection.
• duration - the tooltip says "Duration in seconds (3 or 5)".
• fps - default 24.
• steps - default 20, range 8–50. Distilled likes ~8; dev wants more.
• guidance_scale - default 4.
• generate_audio - default true. LTX's party trick is synced audio; flip off to save cost for silent clips.
• images - optional IMAGE socket for image-to-video.
• loras - CIVITAI_LORAS socket; wire a Civitai LoRA Selector.
• diffusion_model - a CIVITAI_AIR override that swaps the transformer for a community fine-tune (the tooltip names SulphurAI/Sulphur-2-base) while keeping the CLIPs, VAEs, and upscale-LoRA behavior.
• quantity - generate 1–10 videos in one job; each uses a distinct seed (Seed + slotIndex) and re-runs the Comfy workflow.
• api_config - auth override.

Outputs: video, workflow_id, raw_json.

Installing it

ComfyUI Manager → Custom Nodes Manager → "Civitai Comfy Nodes" (publisher civitai) → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

Only requests; no model files.

Common issues

• Auth & Buzz. Civitai account with Buzz; CIVITAI_API_TOKEN on headless installs, Auth node or sidebar otherwise.
• Distilled vs dev. Pick per run. For drafts and quick iteration, distilled at low steps; for quality, dev with more steps. Each choice shifts both time and cost.
• quantity is a spend multiplier. 10 videos in one job = 10 billable generations. Use quantity for seed-sweeping deliberately.
• Off-grid dimensions. Keep width/height divisible by 32 (1280×720 and 1216×704 both qualify) to avoid server errors.
• Early preview. Pack behavior may change without notice; pin the version if you depend on it.