VELVET VICE MiniMax H3 — Image Memory Checkpoint
A RAM/VRAM breadcrumb between the nodes that eat your card
- image
- image
H3 is a memory-hungry model to the point that diagnosing a "where did my VRAM go" mystery is a routine part of running it. VelvetViceMiniMaxH3ImageMemoryCheckpoint is a debugging instrument disguised as a node: it passes an image through untouched while logging a RAM/VRAM snapshot at that exact point in the graph. Drop a few of these around a workflow and you get a memory profile of the whole run instead of guessing.
It's an internal node from the velvet-vice-minimax-h3 pack, and if the design looks familiar, that's because it's a MiniMax H3-branded descendant of a similar checkpoint in the author's LTX video suite. The pattern - a pass-through node that logs system state - transfers cleanly between video model packs because the memory pressure problem is the same shape.
How it works
Two inputs, one output:
image(IMAGE) - the frame batch, forwarded unchanged.checkpoint_label(STRING) - your name for this checkpoint, e.g.after_audio_vae. It's not cosmetic; the label is what makes the console logs readable, because a raw memory dump doesn't tell you which stage of the graph it came from.
Output: image (IMAGE), bit-identical to the input.
Under the hood it calls the pack's memory-snapshot service and prints where you are - RAM and VRAM - to the ComfyUI console. It never modifies the image and never changes execution order, which is the whole contract of a checkpoint node: it's invisible to the render and only visible in your logs.
When to reach for it
Beginners mostly shouldn't need this. The reference workflow is already instrumented, and the pack's System Check and monitors surface the big warnings before a render starts. You reach for checkpoints when things are already going wrong and you need data:
- OOM right at the VAE decode, and you want to know whether memory crept up gradually or jumped at one node.
- Comparing backend choices - does the GGUF path really hold more headroom than native at the same graph position?
- Tuning the live-preview tier or Quality Refine, where a second pass can silently change your peak.
Wire one in before the suspect stage and one after, label them, run, and read the two console lines. The delta is your answer.
Install
Part of the pack:
cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-minimax-h3
then restart. No dependencies beyond ComfyUI's own environment.
One honest caveat: this node reports a snapshot at the moment it runs, and ComfyUI's execution order isn't always the visual top-to-bottom order you expect. If the numbers look odd, remember the checkpoint fires when the executor reaches it, not when you placed it. And it measures your whole process's memory - other things sharing the GPU (a browser, a second ComfyUI, Ollama) will show up in the numbers. The barrier node in this pack exists precisely because of that shared-VRAM reality.
Inputs (2)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | — | |
| checkpoint_label | STRING | image checkpoint | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | — |