🔐 Conduit Seal (Deterministic Cache)

Conduit Seal is the pack's caching planner - the node that turns on deterministic caching and defines its rules. "Deterministic" is doing a lot of work here: the goal is that if you ask for the exact same generation again, you get the exact same result back instantly, verified by checksums rather than trusted blindly. That's a real pain point for people doing reproducible runs, seed farming on locked configs, or building pipelines where identical inputs should mean identical outputs.

How it works

Seal is a settings node. configure_cache reads your four toggles and returns a CONDUIT_SEAL config object describing the cache policy: whether embedding caching is on, whether latent caching is on, how often to snapshot partial steps (step_cache_interval, default 10), and whether checksums get verified (verify_checksums, on by default). It also reports current cache stats and a default cache directory under ~/.conduit_cache.

The honest part is in the details. The real caching in this pack happens inside the functional nodes - ConduitEmbeddingCache and ConduitLatentCache - which use their own in-memory global store (ConduitGlobalCache), not the class-level caches Seal reports on. So the stats Seal shows can sit at zero even while the actual caches are full of hits. And nothing is persisted to that cache_dir yet: both caches live in RAM for the lifetime of the ComfyUI process, keyed by hashes of the prompt (for embeddings) or the full sampling config (for latents). The step-cache described in the source is a stated design, not a shipped feature.

Inputs that matter

• enable_embedding_cache (default true) and enable_latent_cache (default true) - which caches you want the pack using.
• step_cache_interval (default 10, 0–50) - how often to checkpoint mid-sampling if partial-step caching lands. 0 effectively disables it.
• verify_checksums (default true) - the "deterministic" half: recompute hashes and refuse stale entries.
• cache_dir (optional) - where a future disk-backed cache would live.

Output

A single cache_config of type CONDUIT_SEAL, intended for ConduitApply.

Install

Same as every Conduit node:

cd ComfyUI/custom_nodes
git clone https://github.com/joe002/comfyui-conduit-optimizer

Or ComfyUI Manager → search "comfyui-conduit-optimizer" → restart. No downloads, torch 2.0+ only.

Common issues

The big mismatch is between the "checksum-verified deterministic cache" branding and what's live: the actual caches are in-memory and unverified by default paths. If you restart ComfyUI, your "cache" is gone - that's expected. And if you swap checkpoints without clearing caches, the embedding cache can serve embeddings produced by a different CLIP, since its key is just the prompt text. When you change models, either tick force_recompute on the cache nodes or run ConduitCacheClear. For real reproducibility, remember the deeper rule the whole approach leans on: fixed seed, batch size 1, one variable changed at a time - caching is only as deterministic as the underlying sampler.