LTX 2.3 Workflow: All Six ComfyUI Templates

Searching for an "LTX 2.3 workflow" mostly turns up JSON files on Patreon, Discord, and random Google Drive links — each one claiming to be the workflow you need.

You almost certainly do not need any of them.

ComfyUI ships six native LTX 2.3 workflow templates built into the application, and Lightricks' own Python package ships eleven pipelines on top of that. Between them, every common task — text to video, image to video, first/last frame interpolation, lip-synced talking video, structural control, identity personalization — is already covered by something official and maintained.

This guide maps all of them: what each workflow does, which checkpoint it requires, and how to choose. Everything here is sourced from the official ComfyUI LTX-2.3 tutorial and the LTX-2 pipeline documentation, checked July 2026.

The thing most workflow posts get wrong

Third-party workflow bundles are usually built for one release, then never updated. When Lightricks ships a point release — distilled 1.0 → 1.1, upscaler x2-1.0 → x2-1.1 — the bundle silently points at a file you no longer have.

The native templates do not have that problem, and there is a second reason to prefer them: no custom nodes are required. ComfyUI's documentation is explicit that all LTX-2.3 workflows are built in. If a tutorial opens with "first install this node pack," it predates native support.

So the real question is not "where do I download a workflow." It is "which of the six do I open."

The six native ComfyUI templates

Notice the pattern in the right column: four workflows run on dev, two run on distilled, and it is not arbitrary. FLF2V and IC-LoRA are built on the distilled model because those pipelines are constrained by conditioning rather than by guidance — the CLI documentation makes the same split, noting that ICLoraPipeline can only be used with a distilled model.

If you load the wrong checkpoint into one of these templates, it will often still run and produce something. It just will not produce what the template was tuned for. That is the most expensive silent failure in the whole LTX 2.3 stack.

Extra files certain workflows need

Two of the six pull in assets the others do not:

• IC-LoRA Union Control needs ltx-2.3-22b-ic-lora-union-control-ref0.5.safetensors in models/loras/ and moge_2_vitl_normal_fp16.safetensors in models/geometry_estimation/ — the geometry estimator is what turns your reference clip into depth/normal guidance.
• ID-LoRA needs ltx-2.3-id-lora-talkvid-3k.safetensors.

Everything else shares the common set: the FP8 checkpoints, the Gemma text encoder (gemma_3_12B_it_fp4_mixed.safetensors in models/text_encoders/), the distilled dynamic LoRA, and the spatial upscaler in models/latent_upscale_models/.

Choosing between them, by intent

"I have a prompt and nothing else." → T2V. This is the workflow that shows you the model's raw prompt adherence, which is also the fastest way to learn how much detail LTX 2.3 rewards.

"I have an image I want to move." → I2V. Best for maintaining a specific look you already approved — a still you generated elsewhere, a product shot, a character reference.

"I have a start frame and an end frame." → FLF2V. This is interpolation, not animation: you are asking the model to invent a plausible path between two fixed states. Excellent for transitions, logo reveals, and morph-style shots.

"I have a face and a voice track." → IA2V. The lip-sync workflow. Audio quality in equals lip-sync quality out — this is the one place where the input you spend the least time on determines the result the most.

"I want the camera and pose from an existing clip, but a new look." → IC-LoRA Union Control. Structural guidance via depth, pose, or edges. This is the closest LTX 2.3 gets to ControlNet-style direction.

"I want the same person across many clips." → ID-LoRA. Identity personalization with voice cloning.

Want to test the model before you build a workflow? Wiring nodes is worth it when you need repeatability — but not when you are still deciding whether LTX 2.3 fits your project at all. ltx23.app runs it in the browser so you can validate the idea first, then come back and build the pipeline properly.

The CLI side: eleven pipelines

If you run Lightricks' Python package instead of ComfyUI, the mapping is broader. The official pipeline list covers:

1. TI2VidTwoStagesPipeline — the recommended production path: stage 1 at low resolution, stage 2 upsamples 2x with distilled-LoRA refinement
2. TI2VidTwoStagesHQPipeline — same structure, res_2s second-order sampler, typically fewer steps for comparable quality
3. TI2VidOneStagePipeline — single stage, explicitly labeled as primarily educational, typically 512x768
4. DistilledPipeline — fastest: 8 predefined sigmas, 8 steps in stage 1 and 4 in stage 2, no guidance needed
5. ICLoraPipeline — video-to-video and controlled image-to-video (distilled only)
6. KeyframeInterpolationPipeline — uses guiding latents rather than replacing them, for smoother transitions
7. A2VidPipelineTwoStage — audio-driven video; the original waveform is passed through to preserve fidelity
8. RetakePipeline — regenerate only a [start_time, end_time] region of an existing video, with independent regenerate_video / regenerate_audio control
9. HDRICLoraPipeline — linear HDR float output via LogC3 inverse decode, for EXR export
10. LipDubPipeline — lip dubbing matched to a reference clip's speaker identity
11. T2AOneStagePipeline — audio only, no video branch at all

Two of these have no ComfyUI template equivalent and are genuinely underused: RetakePipeline, which fixes one bad second instead of rerolling the whole clip, and HDRICLoraPipeline, which is the only path to gradeable linear-float output.

Retake has hard input constraints worth knowing before you try it: the source video's frame count must satisfy 8k+1 (97, 193, …) and its resolution must be a multiple of 32.

The workflow habit that matters more than the template

Whichever route you take, the structural advice from the official optimization docs is the same, and it is a workflow decision rather than a settings tweak:

• Two-stage by default. Generate low, then upsample. The recommended production pipeline is explicitly the two-stage one.
• Distilled for iteration, dev for the final. 8+4 steps is a different creative loop from 40 steps. Explore cheap, commit expensive.
• Gradient estimation to cut steps. The docs put the range at 20–30 steps instead of 40 while maintaining quality.
• torch.compile is opt-in. --compile mode=reduce-overhead captures CUDA graphs and is described as the main latency lever for the denoising loop — but it reserves static memory pools, so it trades VRAM for speed.

Teams that get consistent output from LTX 2.3 are not using a secret workflow file. They are running the same two-stage template every time and varying only the prompt.

FAQ

Where do I find the LTX 2.3 workflows in ComfyUI? Workflow Templates, inside a current ComfyUI build. They are built in — you do not download them.

Do I need Kijai's nodes? Not for the native workflows. Community node packs exist and are useful for quantized/GGUF experimentation, but the six official templates need no custom nodes.

Which workflow is best for low VRAM? The distilled-based ones (FLF2V, IC-LoRA) plus FP8 checkpoints, at reduced resolution and frame count. In the CLI, DistilledPipeline is the fast path.

Can I use a GGUF checkpoint in these templates? The official templates are built around the FP8 and BF16 builds. GGUF is a community route with its own loader nodes.

What is the difference between FLF2V and KeyframeInterpolation? Same intent, different stacks: FLF2V is the ComfyUI template, KeyframeInterpolationPipeline is the Python equivalent, which uses additive guiding latents for smoother transitions.

Bottom line

The LTX 2.3 workflow question has a boring answer: the official templates are the workflow. Six in ComfyUI, eleven pipelines in the Python package, all maintained against the current checkpoints.

Pick by intent, match the checkpoint the template expects, and put your effort into the prompt instead of the node graph. And when you want a result without opening ComfyUI at all, ltx23.app is the shortcut.

Sources

Verified July 2026 against primary documentation:

1. ComfyUI official LTX-2.3 tutorial — the six native templates, model files, folder paths
2. LTX-2 available pipelines documentation — all eleven pipelines and their constraints
3. LTX-2 pipeline selection guide — official selection logic
4. Lightricks/LTX-2 — official GitHub repository — pipeline list, optimization tips
5. LTX-2 optimization documentation — torch.compile, gradient estimation, FP8
6. LTX-2 installation & CLI flags — runnable pipeline modules
7. Lightricks/LTX-2.3 — Hugging Face repository — checkpoint and upscaler files
8. LTX-2.3-22b-IC-LoRA-Union-Control — structural control LoRA
9. LTX-2.3-22b-IC-LoRA-LipDub — lip dub LoRA
10. ComfyUI-LTXVideo — official integration repository — Lightricks' ComfyUI integration
11. LTX-2 technical paper (arXiv 2601.03233) — model architecture

Template lists change with ComfyUI releases and Lightricks point releases; check the official tutorial page before assuming a template exists in your build.