ModelSamplingContinuousV

This is a niche of a niche, so let's be direct: you don't start a project with ModelSamplingContinuousV. You add it because a workflow you downloaded - usually something video-related, or a model whose training notes mention "continuous v-prediction sampling" - has it sitting between the loader and the sampler, and you want to know what it's doing and whether you can touch it.

The name is the description. It patches your model with continuous-time v-prediction sampling, which is a different math family from the discrete 1000-step schedules that SD 1.5 and SDXL use. In a discrete schedule, noise is defined on a fixed ladder of 1000 steps and the model looks up the nearest rung. A continuous schedule instead treats sigma as a smooth function of a continuous timestep - here with an arctangent mapping: timestep = sigma.atan() / π * 2 and back again via sigma = tan(timestep * π / 2). That tangent-style mapping is what makes the name distinctive; the whole family is the "continuous" counterpart to the older discrete nodes.

The inputs that exist

There are only three, and the sampling dropdown has exactly one choice - v_prediction - so it's really two dials:

• sigma_max (default 500) and sigma_min (default 0.03) - the endpoints of the noise range the schedule spans. The log-sigma grid between them is what the sampler draws from. The defaults are the model-family standard; you'd change them only to reproduce a specific training schedule.

Output is the patched MODEL, same as every model/patch node: clone in, MODEL out, wire it to the sampler.

The honest positioning

This node has effectively zero footprint in the community - you'll find almost nothing about it in real workflow discussions, which is why the "if a workflow tells you to use it" framing is the right one. What you should take away is the family logic: when a model wants continuous sampling, the discrete nodes are quietly wrong in ways that are hard to diagnose, and this node (or its sibling ModelSamplingContinuousEDM, which offers more sampling modes) is how you set it right. If your output looks structurally broken - not stylistically off, but wrong - and your model is a v-prediction continuous sampler, this is the patch that fixes it. Otherwise, leave it alone and save yourself the rabbit hole.