VELVET VICE LTX — Ollama RAM/VRAM Release (Legacy)

The name undersells it, which is rare for this pack. VelvetViceOllamaRelease doesn't generate anything, call an API, or need a key - its entire job is to kick an Ollama model out of memory before LTX rendering starts, because the prompt-writing LLM and the video model are fighting over the same RAM and VRAM. It's also marked Legacy, so before you add it to a fresh graph, know that you're looking at history.

Why this job exists

Velvet Vice's LTX workflows use a local LLM through Ollama as the prompt writer - the standard pattern where an 8B-ish uncensored model turns your rough idea into the long, specific prose LTX actually rewards. The catch is that after the LLM answers, it stays resident, holding its weights in memory. LTX 2.3's BF16 base plus a Gemma text encoder is already a memory monster; a loaded Qwen taking up several more GB on top is how encodes start OOMing. This node is the "please go away now" call before the heavy render.

How it works

It talks to your local Ollama server (default http://127.0.0.1:11434 - no cloud involved). To unload a model it sends an empty /api/generate request with keep_alive: 0 (Ollama's "unload after this request" flag), then polls /api/ps until the model is actually gone, with a timeout_seconds deadline. Input text passes through unchanged - this is a pass-through node with a side effect, not a text transformer.

Inputs and outputs

• text - the string that passes through (and the reason the node can sit inline in the prompt chain).
• enabled - set false and it returns the text without touching Ollama at all.
• server_url - defaults to http://127.0.0.1:11434; change it only if Ollama runs elsewhere.
• model - the one field you actually have to fix. The default value is a specific Ollama model tag; whatever your Ollama install has, run ollama list and type your real model name here. A name that isn't installed will just make the node wait for a model that never unloads.
• timeout_seconds (20) - how long to poll for the unload.
• strict_cleanup - true means a failed release raises an error and halts the run; false downgrades to a warning.

Output is text - same string you put in.

Installing it

Part of the Velvet Vice LTX pack. ComfyUI Manager → search "VELVET VICE - LTX" (registry velvet-vice-ltx), or:

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-ltx

Restart ComfyUI and Ctrl+F5. The pack itself has zero Python dependencies, but this node is useless without a running Ollama server (ollama serve) on the same machine. One gotcha from the README: if you upgrade an older Velvet Vice install, never merge the old custom-node directory into the new one - install fresh.

Common issues

• It errors saying it can't reach Ollama - ollama serve isn't running, or server_url points somewhere wrong. This is the most common failure and it's almost always "Ollama isn't running."
• It times out waiting for the unload - the model name is wrong or the server is busy generating for someone else. Fix the model tag or bump the timeout.
• Why am I in Legacy? The current barrier node (VelvetViceOllamaReleaseBarrier) does the same job but knows which models the Prompt Director actually used, instead of you hardcoding one name. Use that for new graphs; keep this only if you have V1.1 workflows that reference it by ID.

It's a boring, single-purpose node, and that's fine. The interesting part is that it's being kept around purely so old workflows don't break - which is more care than most packs give their legacy.