H3RelayInternalMemoryRelease

A node with no inputs and one string output, marked as an output node, that you will never drag onto a canvas yourself. That's Memory Release, and its entire job is summed up by the README: an explicit release barrier that clears ComfyUI's executor cache and loaded models between jobs. It's the punctuation at the end of a stage - the thing that makes sure a finished H3 shot doesn't leave the LTX models resident in VRAM when the next stage is trying to load RIFE.

Why a no-op-looking node matters

ComfyUI keeps loaded models in an executor cache because reloading them is expensive, and for most workflows that caching is pure win. For a long staged pipeline it's a liability: queue five H3 shots, five LTX finishes, and five interpolations as one monolithic prompt and every model family stays loaded at once. That's how a 33B H3 model, a 22B LTX transformer, the Gemma text encoder, and the RIFE checkpoint end up competing for the same 24GB.

The staged execution path solves that by splitting the graph into per-stage prompts, and Memory Release is the explicit barrier between them: it unloads models and clears the executor cache so each stage starts with a clean slate. The trade-off is a reload cost between stages - the README is upfront that the default minimum-RAM release barrier "trades a small reload cost for predictable memory."

There's a nuance worth appreciating: cached model patchers stay lazy, so a cache hit doesn't materialize checkpoint weights on the GPU. The release barrier isn't the only memory defense in the pack - it's the explicit, deterministic one.

Where it shows up

When you use the Run staged · bounded RAM action on H3 Relay · Assemble, the staged orchestrator builds each partial prompt to construct only the model family its target requires, restores accepted state from disk manifests, and inserts this release barrier between jobs. A failed stage stops the sequence instead of queueing invalid dependents. So this tiny node is doing real work every time you run a long sequence, even though you'll never configure it.

The output is a single status string - the barrier confirms it ran. That's the whole interface.

The takeaway

If you see Memory Release in an expanded graph, leave it alone. If you're running the plain monolithic queue instead of the staged action, you won't see it at all - and that's exactly when you might want the staged action, because it's the difference between a multi-shot sequence that fits in VRAM and one that OOMs at shot three. For anything longer than a couple of shots on a 24GB card, the staged path is the one to reach for.

Install: ComfyUI Manager search H3 Relay, or git clone https://github.com/akatz-ai/h3-relay.git into custom_nodes. ComfyUI 0.32.0+, FFmpeg on PATH, and the model files from MODELS.md are the real requirements; no extra Python dependencies. The MiniMax H3 community license geofence (US/EU/UK/Korea) applies on the H3 side regardless.