Cloud LTXV Scheduler

LTX-2 is a flow-matching model, which means the old "pick a scheduler and pray" game is mostly over - what matters instead is a set of shift parameters that control how sampling effort splits between composition and detail. Cloud LTXV Scheduler is the node that generates those sigmas, contributing an LTXVScheduler step to the cloud workflow JSON and returning a CLOUD_SIGMAS handle. If you're building a "sampling primitives" style chain (model → guider → scheduler → noise → custom sampler), this is the scheduler slot.

Flow-matching shifted the tuning knobs: instead of Karras vs. exponential, you get shift values, and this node exposes LTX's two-shift variant. The defaults - max_shift 2.05, base_shift 0.95 - are the values LTX ships with and most workflows never touch, but knowing what they do beats randomly wiggling them.

Inputs

• steps - INT, default 20. Number of sampling steps.
• max_shift - FLOAT, default 2.05. The shift at the "compositional" end of the trajectory.
• base_shift - FLOAT, default 0.95. The shift at the "detail" end.
• stretch - BOOLEAN, default true. Whether to stretch the schedule to the terminal sigma.
• terminal - FLOAT, default 0.1. The final sigma value the schedule lands on.
• latent (optional) - a CLOUD_LATENT. Some LTX variants shape the schedule from the latent's dimensions, so the node accepts one when you have it.

Output: sigmas (CLOUD_SIGMAS), which feeds the sampler's sigmas input in a Cloud Sampler Custom Advanced chain.

Installation

Same pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Dobidop/ComfyUI-CloudAPI-worker

Copy config.json.example to config.json, paste an API key from https://platform.comfy.org/profile/api-keys, restart. Manager-installable as "ComfyUI-CloudAPI-worker"; deps: requests, Pillow, safetensors.

Gotchas

• CLOUD_SIGMAS is its own handle type - it plugs into the cloud custom sampler, not into nodes expecting a local sigmas tensor. Don't try to swap it into local sampler nodes.
• If you're not chasing quality tweaks, leave the shifts at default. The KB's flow-matching coverage is blunt that shift tuning is model-specific and that wrong schedules are how flow models get mangled - the defaults exist because they're the known-good values.
• Proof-of-concept pack; the LTX nodes were added in an update the author called working "well with the example workflow at least." The example LTX workflow shows the intended sampler chain (scheduler → custom sampler with guider and noise) - mirror it.