JoyLTX Keyframes (end / mid frames)

A text prompt says a lot, but it can't say what the last frame of a take should look like. JoyLTX Keyframes is how you pin that down on the Take canvas: feed it an end_image and the take's final frame will be the one you chose. It also takes mid_images at specific frame indices, and a ref_image of a person for soft identity anchoring. The clever bit is how - nothing gets composited or pasted. The images become LTX-2.5 keyframe tokens (AddGuide) that condition the generation and are cropped out after pass 1. They steer; they never appear in the output.

You wire your existing positive/negative CONDITIONING, vae, and latent straight through, and get the same three back out with guides attached. That's the whole interface: the node sits in the conditioning path and modifies what flows past. The required strength (default 0.8) controls how hard the keyframes pull - 1.0 reproduces the frame, lower gives the model room. mid_frame_indices is a comma list of pixel-frame indices, one per mid image, like 48,96; they're rounded down to the 8n latent grid, which matters because LTX's temporal structure works in 8-frame chunks. mid_images is the batch of mid keyframes in that same order. The optional ref_image is different in kind: it's attached at frame 0 at ref_strength (default 0.85) as a soft in-context reference - the person in the photo is the person in the take, but framing stays free. That's the same mechanism Refs by Name uses, exposed for a single photo.

Where people get burned: the crop. These are appended tokens, and the pack's canvas includes an LTXVCropGuides node after pass 1 that removes them before the upscale/decode. Leave it in place. Remove it and your keyframes will leak into the final frames. The node itself prints how many keyframes it attached and reminds you to crop them after pass 1.

If you connect none of the optional images, the node passes everything through unchanged at no cost - it's a zero-overhead member of the conditioning chain until you use it. That's the right design for a canvas where the end-frame gate and the reference gate are toggles you flip per project.

On the Multishot side, the equivalent is the sampler's end_images slot (one keyframe per shot at each shot's last frame) - JoyLTX Keyframes is the Take-canvas version, for a single take. Install is just the JoyLTX25 pack; there are no extra models. If you want a take to land on a specific last frame, or a character to look like your reference photo, this is the node.