SSG Smart Vault

SSG Smart Vault is the suite's memory node: an inline buffer that records a run's 24 lanes into RAM and can play them back later without re-running whatever produced them. The pitch is prompt iteration and upscaling - you generate a base latent once, "record" it, then iterate on the downstream upscale/refine stages over and over while the Vault hands back the cached tensors instead of re-running the whole model. If that sounds like the ideal workflow-caching node, it half-delivers, and the half it misses matters.

The state machine is a tri-state mutex you drive with two booleans, flush_switch (default on) and cache_switch (default off):

• BUFFER (recording): flush_switch on, cache_switch off. Signals pass through live and are written into the in-process cache. This is the "warm the vault" pass.
• PLAYBACK: cache_switch on. The node ignores its live inputs entirely and returns whatever's in the cache for that vault. This is the fast iteration mode.
• FROZEN: both off. Signals pass through live but nothing is written, so the stored cache stays as it was.

The frontend renders a state badge on the node title (❄ / 🔴 / ⚡) so you can tell at a glance. Inputs and outputs are the usual 24 wildcard SSG_0–SSG_23 bus, and each Vault is keyed by a vault_id the frontend auto-generates - distinct Vaults don't stomp each other's cache.

Now the reality check, because the README oversells. It calls PLAYBACK "severed" and promises "zero upstream execution time" - that's not how ComfyUI works. The Vault's Python just reads the cache and returns it; the nodes wired into its inputs will still execute, because ComfyUI schedules by wired dependencies and doesn't know the Vault is ignoring them. You get the caching win - downstream iterating on cached tensors is genuinely fast - but you don't get ComfyUI magically skipping your model loaders. The second limitation is that the cache lives in process memory (torch._ssg_vault_registry). Restart ComfyUI and the vault is empty; playback on a cold vault returns Nones until you re-record. Plan around both and it's a handy iteration tool; expect either and you'll be confused.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/SgtSauv/ComfyUI-SSG-Smart-Suite.git

Restart, or install via Manager by searching "SSG Smart Suite". No model files, no Python dependencies (requirements.txt literally says "No external dependencies"), just a console boot banner on startup.

Common issues

• Playback returns Nones → cold vault: the cache was never recorded in this process, or ComfyUI was restarted. Run once in BUFFER, then switch to playback.
• "Severed" doesn't actually skip upstream work → correct; the Vault ignores its inputs in playback but ComfyUI still runs what's wired in. Bypass the expensive loaders yourself if that's the goal.
• Thin community → this suite has almost no Reddit presence as of 2026; GitHub issues is your support channel.