LTX 2.5 Sigma Preset

If you've done any LTX enhancement work - the face-fix and video-enhance pipelines where you push a clip back through the model at low denoise to clean it up - you know the pain of tuning sigmas. LTX 2.5 Sigma Preset replaces "type four magic numbers into a string" with a dropdown of human labels. It's a tiny node, and it's supposed to be.

This is a preset generator, not a sampler. You pick a strength, and it hands you a ready-made SIGMAS tensor that the rest of your sampling graph consumes. It exists because the VRGameDevGirl enhancement workflows (the ones that repair faces or restore a whole video through LTX) need a schedule that reaches 0.0 at the end so the pass runs to clean output - and because most people who use those workflows shouldn't be hand-rolling sigma lists on a deadline.

What the presets actually mean

The schedule is four numbers, and the first one does almost all the talking:

| Preset | First sigma | Vibe | |---|---|---| | Enhance Only - Preserve Original | 0.30 | Smallest change; near-identity refresh | | Very Low - Subtle Detail Recovery | 0.50 | Shake off compression, keep the frame | | Low - Light Enhancement | 0.65 | | | Medium - Balanced Enhancement | 0.75 | | | High - Noticeable Refinement | 0.85 | | | Very High - Strong Reinterpretation | 0.95 | LTX gets room to redraw lighting/texture/motion |

The first sigma sets how much noise the pass starts from - in img2img terms, that's your denoise strength. Lower = "keep my video, just polish it"; higher = "LTX may do its own thing." The author's own warning in the node description is worth taking at face value: Enhance Only minimizes changes but cannot guarantee perfect preservation. LTX can still move faces, change lighting, shift textures while denoising even at 0.30. If you're repairing a specific face you care about, you verify the result, you don't assume "Enhance Only" is a no-op.

Inputs and outputs

• preset - the seven-way dropdown above, plus Custom - Enter Your Own Values.
• custom_sigmas - only read when Custom is selected. Comma-separated numbers, e.g. 0.85, 0.725, 0.4219, 0.0. The node validates them (at least two, finite, no empty entries) so a typo raises a readable error instead of silently corrupting the schedule.
• sigma_values - a read-only preview of whichever schedule is currently active. Handy for seeing what a preset resolves to before you queue.

The single output, sigmas (type SIGMAS), plugs into a sampler's sigma input - in the enhancement graphs it feeds the KSampler-style node that re-samples your video latent through LTX. Output of the node is a plain one-dimensional torch tensor, which is exactly the convention ComfyUI's sampler expects, so it drops straight in.

Install

It ships in the VRGameDevGirl pack - one install gets you this and the whole video-builder suite:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git

Or find the pack by name in ComfyUI Manager. Restart, then hard-refresh the browser so the new UI files load. Expect a heavier pip install than most one-node packs: the requirements include kornia, librosa, transformers, voxcpm and llama-cpp-python (the README tells Windows users to install Cython scikit-build-core first, and prefers Python 3.12 over 3.13 on portable builds). Also note the builder guide's standing advice: workflow templates don't download models - the actual LTX 2.5 checkpoint and its VAE are separate downloads you pick in Builder Settings. A sigma preset can't save you from missing model files.

The trap to avoid

The Custom option is there for people who know exactly what they want, and the number one mistake is assuming a higher custom starting sigma on "Enhance Only" just means "more enhancement." Past ~0.5 the pass stops enhancing and starts regenerating - which is why Very High is honestly labeled "Strong Reinterpretation." If a scene comes back looking like a different video, you didn't find a bug; you found the 0.85 preset doing its job.