LTX Forward Model Sampling Pred (Ray)

LTX Forward Model Sampling Pred (Ray) is one of the two "direction" nodes that make LTX video extension work inside Raylight. If you've seen the LTX community's favorite trick - generate a short clip, then keep going past the end, or pre-pend frames before the start - these are the nodes that flip the sampler into the right mode for it.

Here's the mechanism. LTX's flow-matching model predicts the direction of denoising. For normal text-to-video you sample from noise toward an image. For extension you want the opposite maneuver: take an existing (already clean) latent and run the sampler backward through the noise schedule, generating the frames that would have led to it. ComfyUI implements this by swapping the model's sampling object for one where the "input noise scaling" is inverted - calculate_input returns noise untouched, calculate_denoised passes the model output through - while inverse_noise_scaling does the reverse-domain math (latent / (1 - sigma) for the reverse flavor).

This node does exactly that swap, on every Ray worker: it clones the model and patches in a ModelSamplingFlux-based sampling object with a shift of 1.15 and an InverseCONST prediction mode. That shift of 1.15 is the LTX family's standard - same value its sampling node uses.

When you reach for it

You use this in a forward extension workflow: you have a finished clip's last frames, and you want LTX to generate what happens next. The counterpart node, RayLTXReverseModelSamplingPred, does the backward direction (pre-pending frames before the start). They're a matched pair - pick the one matching the direction you're extending, wire ray_actors through it into the Ray sampler, and keep the frame-count arithmetic from the main LTX workflow intact (frame count = 8n+1).

The catch

These nodes have no tunables - just ray_actors in, ray_actors out. That's a feature: there are no settings to get wrong, and the moment you add one of these to the graph it's active. The flip side is you have to know why it's in the workflow. Drop it into a normal T2V graph out of curiosity and you'll change sampling behavior for no benefit. Use it for its intended job - extending - and it's a zero-config requirement rather than a setting to tune.

Install is the pack standard (ComfyUI Manager → "raylight", or clone + pip install -r requirements.txt), no model downloads. Works with the pack's LTX-2/2.3 support, which the README lists as USP-supported.