This guide covers running LTX-2 in ComfyUI on Runpod. LTX-2 is Lightricks' open-weights audio-video model, and its distinguishing feature is that it generates both tracks in a single pass: motion, dialogue, sound effects, and music come out of the same forward pass rather than being stitched together afterward.
By the end you'll have a pod running ComfyUI with an LTX-2 workflow loaded and your first synchronized clip generated. This guide covers which checkpoint to pick for your VRAM budget, how the workflow templates differ, and how to write prompts the model actually follows.
What is LTX-2?
LTX-2 is a 19-billion-parameter DiT-based audio-video foundation model from Lightricks. Lightricks announced it in October 2025 and released the full weights, inference pipelines, and training code to the open-source community that November.
Most open video models generate silent frames and leave audio to a separate pipeline. LTX-2 produces both together, which is why lip movement lines up with speech and impact sounds land on the frame where the impact happens.
The model ships in several checkpoints:
- ltx-2-19b-dev: Full model in bf16. The most flexible option and the one to fine-tune against.
- ltx-2-19b-dev-fp8: The same model quantized to fp8. Roughly half the memory footprint.
- ltx-2-19b-distilled: Distilled to eight steps at CFG 1. Substantially faster, and the right choice for iteration.
- ltx-2-spatial-upscaler-x2-1.0: 2x spatial upscaler for higher output resolution.
- ltx-2-temporal-upscaler-x2-1.0: 2x temporal upscaler for higher frame rate.
Capability-wise it covers text-to-video, image-to-video, and video-to-video, with Canny, Depth, and Pose structural control through IC-LoRAs, keyframe interpolation between reference images, and automatic prompt enhancement.
Lightricks has continued to iterate on the LTX line since this release. If you want the newest checkpoint rather than LTX-2 specifically, check the ComfyUI video templates and Lightricks' Hugging Face repositories for current versions. The deployment path in this guide is the same either way.
What is ComfyUI?
ComfyUI is an open-source, node-based interface for generative AI pipelines. Instead of writing inference code, you connect nodes, each one a loader, sampler, conditioning step, or output, into a workflow graph.
LTX-2 is natively supported in ComfyUI, and the audio-video workflows are non-trivial graphs: separate audio and video VAEs, latent concatenation, and upscaler passes. Running them from ComfyUI's template library means you don't wire any of that yourself.
Why run LTX-2 on Runpod?
A 19B video model with synchronized audio generation is not a laptop workload. The bf16 checkpoint wants 32 GB of VRAM or more, and even the fp8 build is tight on a 16 GB card once you push past a few seconds of output. Video generation is also naturally bursty: you need serious hardware for an afternoon of generating, not permanently.
On Runpod you select the GPU tier that matches your checkpoint and resolution target, deploy a ComfyUI template, and pay per second while the pod runs. Nothing accrues while you're not generating, and you can move from a 24 GB card for testing to an 80 GB card for final renders without changing anything else about your setup.
GPU requirements for LTX-2
Video VRAM scales with resolution, clip length, and frame rate together, so treat these as practical starting points rather than hard limits:
- 16 GB VRAM (RTX 4080, A4000): fp8 or distilled checkpoints at 540p, short clips
- 24 GB VRAM (RTX 4090, RTX 3090, L4): fp8 and distilled comfortably at 720p, bf16 with offloading
- 48 GB VRAM (L40S, A6000, RTX 6000 Ada): bf16 at 720p, longer clips, upscaler passes in the same graph
- 80 GB VRAM (A100 80GB, H100): High-resolution output, extended clips, batch generation
Start with the distilled checkpoint regardless of your card. Eight steps at CFG 1 turns a two-minute generation into something you can iterate on, and you can re-render selects on the full model once the prompt is right.
Templates for running LTX-2 on Runpod
ComfyUI (official Runpod template) Runpod's first-party template, running the current ComfyUI build on CUDA 12.8 with ComfyUI-Manager, KJNodes, Civicomfy, and ComfyUI-RunpodDirect pre-installed. ComfyUI copies to your workspace on first boot. Wait for [ComfyUI-Manager] All startup tasks have been completed. in the pod logs before opening the interface. You add the LTX-2 checkpoints yourself, and ComfyUI's workflow templates prompt you for the downloads.
For other options, browse console.runpod.io/hub and search for ComfyUI. The Hub is a live catalog, so check the image version and last-updated date on any community template before deploying.
Step 1: Create a Runpod account
Go to runpod.io and sign up with email, GitHub, or Google. Add credits or a payment method so you can deploy a GPU.
Step 2: Deploy a ComfyUI pod
- Open the Pods section and select your GPU. An RTX 4090 is a reasonable starting point for fp8 and distilled work; step up to an L40S or A100 for bf16 at 720p and above.
- Click Change Template and select the official ComfyUI template.
- Set container disk to at least 80 GB. The 19B checkpoints are large, and you'll likely want more than one.
- Attach a network volume so the checkpoints survive between sessions. Re-downloading tens of gigabytes on every deploy gets old fast.
- Click Deploy.
The pod comes up in under 30 seconds. Template setup takes a few minutes on first boot.
Step 3: Open ComfyUI and load an LTX-2 workflow
When the pod shows a running state, click Connect, then open port 8188.
Update ComfyUI before you load anything. LTX-2 support depends on recent core nodes, and a stale build is the single most common reason a workflow loads with missing nodes.
Then go to Template Library > Video and pick your workflow:
- Text-to-Video or Text-to-Video Distilled for generating from a prompt alone
- Image-to-Video or Image-to-Video Distilled for animating a starting frame
- Depth-to-Video, Canny-to-Video, or Pose-to-Video for structural control from a reference clip
ComfyUI prompts you to download the models the selected workflow needs. Accept, wait for the downloads, and the graph is ready.
Start with a distilled workflow. It gets you to a first output faster, and confirms the environment works before you commit to a long bf16 render.
Step 4: Generate your first clip
Write the prompt as a paragraph, not a tag list. LTX-2 responds to chronological description of what happens. Start directly with the action and stay literal.
Structure it in this order:
- The main action in one sentence
- Specific movements and gestures
- Character and object appearance
- Background and environment
- Camera angle and movement
- Lighting and color
- Any change or sudden event
Keep it under roughly 200 words. Past that, adherence degrades.
Start small. Generate at a lower resolution and short duration first. Confirm the motion and audio are what you wanted, then raise resolution or clip length. Frame count drives memory harder than resolution does, so if you hit an out-of-memory error, cut frames before you cut pixels.
Queue the prompt. Click Queue Prompt and watch the pod's GPU utilization from the Runpod console while it runs. Output lands in the ComfyUI preview and saves to /workspace/ComfyUI/output.
Step 5: Add upscaling passes
Rather than generating at high resolution directly, generate at a comfortable resolution and upscale. LTX-2 ships dedicated upscalers for exactly this:
- The spatial upscaler doubles output resolution
- The temporal upscaler doubles frame rate
Two 2x passes on a 540p 24 FPS base clip cost far less VRAM than generating the equivalent output natively, and the results hold up. This is the standard way to get high-resolution LTX-2 output without an 80 GB card.
Pro tips
Distilled for iteration, dev for finals. Eight-step generation at CFG 1 is fast enough to actually explore prompts. Re-render your selects on the full model once the prompt is locked.
Audio is generated, not attached. If dialogue or sound effects matter, describe them in the prompt. LTX-2 produces them in the same pass as the video, which is why sync holds, but it can only generate what you asked for.
Reserve VRAM in ComfyUI when you're close to the edge. ComfyUI's --reserve-vram startup flag combined with the low-VRAM loader nodes buys you headroom on cards that are marginal for your chosen checkpoint.
Export your workflow JSON. When a graph works, export it. Re-importing on a fresh pod restores the exact configuration, including sampler settings and upscaler chain.
Stop your pod when you're done. Runpod bills by the second while a pod runs. An idle 80 GB card is an expensive way to leave a browser tab open.
Wrapping up
LTX-2 closed a real gap in open video generation: audio that matches the frames instead of being layered on afterward. Running it in ComfyUI on Runpod means you get the full graph, including control LoRAs and upscaler chains, on whatever GPU tier the job actually needs.
Ready to try it? Deploy the ComfyUI template and start with a distilled text-to-video workflow.
FAQ
What GPU do I need to run LTX-2 on Runpod?
An RTX 4090 or another 24 GB card handles the fp8 and distilled checkpoints at 720p. The bf16 checkpoint wants 32 GB or more, so an L40S, A6000, or A100 is the better fit. Sixteen-gigabyte cards work for fp8 at lower resolutions and short clips.
What's the difference between the dev and distilled checkpoints?
Dev is the full model, flexible and trainable, running the standard step count. Distilled is compressed to eight steps at CFG 1, generating several times faster with some quality trade-off. Most people iterate on distilled and re-render finals on dev.
Does LTX-2 really generate audio and video together?
Yes. It's a single audio-video model, not a video model with a separate audio pipeline bolted on. Motion, dialogue, sound effects, and music are produced in the same pass, which is what keeps them in sync.
Can I control the output with a reference video?
Yes. LTX-2 supports video-to-video control through IC-LoRAs for Canny edges, depth maps, and pose. ComfyUI ships a workflow template for each. It also supports keyframe-driven generation, interpolating between reference images.
Can I run LTX-2 as an API instead of a UI?
Yes. Once you have a ComfyUI workflow you're happy with, containerize it and deploy it to Runpod Serverless. Your application posts generation requests to the endpoint, which scales from zero and bills per second of compute rather than per hour of uptime.
