VELVET VICE MiniMax H3 — Ollama Release Barrier

If your H3 workflow uses the Prompt Director's vision mode, the graph is holding an Ollama vision model in VRAM to analyze your source image and write the prompt. Then the actual render needs that same VRAM for a 33B diffusion model, a Qwen3-VL text encoder and two VAEs. Those two things cannot coexist comfortably on one GPU. VelvetViceMiniMaxH3OllamaReleaseBarrier is the node that politely asks Ollama to unload its model before H3 takes the stage.

It sits in the pipe between the Prompt Director and the model-loading stage: prompt analysis happens up front, and this barrier guarantees the vision model is released before the heavy diffusion load begins.

How it works

The node subclasses a shared VelvetViceOllamaReleaseBarrier and calls into the pack's Ollama client service. Its inputs:

• prompt_package (VELVET_VICE_PROMPT_PACKAGE) - the Director's output bundle, so the barrier knows which prompt stage preceded it.
• strict_release (BOOLEAN, default true) - how hard it insists. Strict means the release must actually happen; with strict off, a failed Ollama release won't necessarily block the render.
• timeout_seconds (INT, default 20) - how long it waits for Ollama to confirm the model is gone before giving up.

Output is a single prompt (STRING) - the resolved prompt continues down the chain, because the barrier is a pass-through for the text while being a hard stop for the model.

Why a dedicated barrier instead of just an "unload" call buried in the loader? Ordering. Graph execution isn't guaranteed to happen in the order you'd naively expect, and nothing is worse than H3 loading its weights while a vision LLM is still resident, only to OOM halfway through and take down the whole queue. The barrier makes the sequencing explicit and verifiable.

When it matters

This is one of those nodes that's pure overhead if you never use vision-assisted prompting. Run the Prompt Director in MANUAL mode - you write the prompt - and there's no Ollama model to release, so the barrier is mostly ceremony. Use STANDARD VISION or the ADULT ASSISTED modes and it's doing real work every render, because those paths load an Ollama vision model per session.

Install and troubleshooting

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-minimax-h3

or ComfyUI Manager → velvet-vice-minimax-h3, then restart. Ollama itself is a separate prerequisite you install on your own - the pack just talks to its HTTP API.

The failure mode to know: renders that stall right after the analysis stage are usually a release that never confirmed - Ollama unloaded but the barrier's timeout expired first, or Ollama isn't running at all and strict_release is refusing to proceed. If you see it hang around the 20-second mark, check that Ollama is actually up and that the vision model name matches what the Prompt Director configured. And if you're mixing this pack with the separate "Velvet Vice Zen" H3 package, the two are runtime-isolated by design - a release barrier here doesn't touch models that other package loaded.