LTX Take Controls (Joy-LTX 2.5)

This node is the odd one out in the pack in the best way: it's not for MiniMax-H3 at all. It's the MASTER CONTROLS panel for the pack's other canvas, the LTX-2.5 single-generation one (the "Joy-LTX 2.5" workflow). Where H3 renders short blocks and chains them, LTX renders a whole take in ONE generation - and LTX has a cost model that punishes guessing. H3LTXTakeControls does for LTX what the H3 studio controls do for H3: it sizes the render to your card so you don't find out at step 200 that the take was never going to fit.

How it works

LTX's cost is tokens = (W/32)(H/32) × latent frames, and the upscale pass runs on the output grid. That's the lever this panel pulls: the upscale factor decides the real cost. x2 means 4x the pixels of pass 1, x1.5 means 2.25x, none means pass 1 only. So upscale defaults to auto - the largest output that fits your card - and as you push take_seconds up, auto steps the upscale down (x2 → x1.5 → none) and prints which, keeping you on the edge of what fits instead of over it. The measured comfort zone on a 24GB card is 8s at x2, ~14s at x1.5, at 960x544.

The inputs that matter

• width, height - render (pass-1) size; output = this × the upscale factor, so 960x544 x2 = 1920x1088. Multiples of 32.
• take_seconds - 1 to 60, length of the whole take in one generation.
• beat_seconds - how long each written sentence/beat should be; the writer gets take_seconds / beat_seconds beats to write. This is the pacing knob that keeps dialogue from cramming.
• upscale - auto / x2 (the official LTX-2.5 spatial upscaler, 4x pixels in pass 2) / x1.5 (the LTX-2.3 x1.5 file, which works on 2.5 latents - verified - and is ~1.8x the seconds of x2) / none.
• vram_gb - 0 = read the card. Set a number to plan for another card, useful when you're tuning a workflow you'll hand to someone with a 16GB card.

Outputs: pass1_width, pass1_height, frames, two_pass (whether pass 2 runs), use_x15, beats, beat_frames, and a summary string. Wire these into the LTX sampler, the writer, and the upscaler.

Where it fits

The LTX ecosystem has always been the speed-first corner of local video - LTX's defining trait is fast, short generations. This workflow takes that and gives you a take: a single generation of 8-14 seconds with audio, sized to the card automatically, with the writer's beat count driven by the same panel so dialogue matches the actual length. If you're on the LTX-2.5 path and want "type a length, render a take" without doing token math yourself, this is the node. (It's also proof of how seriously this pack's author documents cost models - the tooltips here are little essays.)

Install

Ships in the H3 Multishot pack:

cd ComfyUI/custom_nodes
git clone https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

or ComfyUI-Manager → H3 Multishot. Restart; ComfyUI v0.30.0+. The LTX workflow also needs the LTX-2.5 model files and ComfyUI's own LTX nodes, which ship with ComfyUI - check the workflow's requirements for the exact upscaler files (the x2 and x1.5 variants are separate downloads).

The trap: vram_gb = 0 reads your card, but if you've got other things loaded, "reads your card" reads the total, not the free headroom. The panel's estimates assume a clean card. If your take OOMs on a busy machine, drop take_seconds or pin upscale to x1.5 - that's the knob auto won't turn for you mid-render.