IAMCCS LTX Release VRAM Before Sampler

LTX-2.3 is a 22B model with audio built in - it's also exactly the model that will OOM you at the worst possible moment unless something cleans house first. In a long V2V run, the stuff you loaded earlier (the text encoder, the VAE that made your guide latents, the conditioning machinery) is still parked in VRAM when the sampler asks for the big diffusion model. ComfyUI's DynamicVRAM will happily juggle models, but it has a known blind spot: it does not automatically evict one dynamic model to make room for another dynamic model. That's the gap this node is built for.

IAMCCS_LTXReleaseVRAMBeforeSampler is a barrier, not a processor. It takes latent and guider in, and at that point in the graph - where your guide latents and conditioning are already computed, so there's nothing left to lose - it calls unload_all_models(), soft-empties the CUDA cache, and passes the latent straight through. The code even prints a [IAMCCS_V2V_VRAM_GUARD] line showing free VRAM before and after, so you can watch it recover a couple of gigabytes in your console. Output is the same latent it received; nothing is transformed, which is why it's so easy to trust.

The placement is everything. You want it after everything that produced the latents and conditioning you still need, and immediately before the LTX sampler - because everything it unloads, you lose. Put it too early and you evict the text encoder while conditioning is still being built. The comment in the source is explicit about the contract: "Both guide latents and conditioning are complete at this barrier." If your graph builds conditioning downstream of this node, you've placed it wrong.

Inputs are just latent (LATENT) and guider (GUIDER) - the guider isn't modified or even read deeply; it's there because in the IAMCCS V2V engine a guider is always in play at this stage and the node's whole job is to sit in that flow. Output is the single latent passthrough.

Installation is pack-standard: search IAMCCS in ComfyUI Manager, or clone https://github.com/IAMCCS/IAMCCS-nodes.git into custom_nodes and restart. It has zero dependencies beyond ComfyUI's model_management - no pip packages, no model files. Its real prerequisite is the graph around it: this is a tool for LTX-2.3-class AV runs, and those want a recent ComfyUI with native LTX audio-video nodes plus your LTX model, text encoder, video VAE and audio VAE (the pack's SuperNodes requirements doc is the checklist).

Where people get confused: this is not a magic VRAM multiplier. On a GPU that's genuinely too small, unloading the VAE gets you through the forward pass but the decode after it will pay the cost in reload time. The win is avoiding the OOM crash and the mid-run restart, not making the model smaller. And it's a deliberate trade: after the barrier, your VAE is cold, so the post-sampler decode re-loads it. If you're on a 12–16GB card and already juggling, pair it with the pack's low-RAM VAE-decode-to-disk nodes instead of expecting this one barrier to carry the whole run. On a 24GB card it's the difference between a 40-second hitch and a hard crash, and it's cheap insurance.