Sync Barrier

If your ComfyUI log shows Unloaded partially repeating between every prompt in a batched multi-prompt workflow, this is the node you were looking for. That log line - and the minutes of wall time it represents - is the symptom: CLIP loads, gets evicted to make room for the UNet, reloads for the next branch, gets evicted again. Sync Barrier fixes the ordering problem underneath it. It's a true execution barrier: its outputs only resolve once every connected input is ready, so ComfyUI finishes one entire phase before the next starts. One CLIP load, one UNet load, one VAE load. Load thrash gone.

The author's numbers make the pitch concrete - a 3-image Z-Image Turbo batch on Apple Silicon:

| Approach | Wall time | |---|---| | 3 separate single-image runs | ~1500s | | Batched, no barrier | ~1286s (repeated Unloaded partially) | | Batched + Sync Barrier (+ Phase Unload) | 887s (14:47) |

That's the difference between "the batch runs and you watch the log churn" and "the batch runs". On a machine with tons of VRAM headroom you may see almost nothing - the pain is concentrated on constrained VRAM and unified-memory setups, which is why this pack's workflows are Z-Image-on-MPS flavored.

How it works

SyncBarrierN is a wildcard (*) passthrough with up to 1000 in_N/out_N socket pairs. The mechanism is the dependency graph itself: because each output depends on all connected inputs, Comfy's executor can't run anything downstream of the barrier until every branch feeding it has resolved. That's the "barrier" - it's not magic, it's just refusing to let phase two start until phase one is completely done. Unconnected optional inputs don't block; only wired ones count.

The fun part is the UI. The Python declares all 1000 slots, but a small JS extension starts each node with a single in_1/out_1 pair and auto-grows sockets as you connect - plug into the last input and the next pair appears, up to the 1000 cap. You never scroll past hundreds of empty pins for a 3-branch batch.

How to use it

1. Duplicate your per-prompt branch (CLIP encode, sampler, etc.) for each prompt.
2. After each phase's branch outputs - e.g. all the CLIP CONDITIONINGs, or all the sampler LATENTs - drop in a Sync Barrier.
3. Wire in_1…in_N from each branch, then pull out_1…out_N into the matching next-phase nodes.
4. On macOS/unified memory, follow the barrier with Phase Unload on every wire into the next phase. The barrier stops the ordering thrash; the unload stops models staying resident (Comfy's eviction treats macOS "available" RAM as free, so it doesn't evict what it should).

Where people get burned

• It delays everything downstream - by design. A UNet barrier waits for every sampler before any VAE decode, which means no PNGs on disk until the whole sample phase is done. If ComfyUI crashes mid-batch, your finished branches were only in-memory latents. Mitigate with per-branch Save Latent (Atomic) before the barrier, then decode later.
• It's not a single-prompt tool. It exists for one graph running several branches; a one-image workflow gains nothing and just adds a node.
• The barrier is not the unload. The README is explicit: one Phase Unload on a single branch doesn't gate the others - put it on every wire into the next phase, and never hang it off an individual save (it can fire mid-phase).
• Wildcard sockets hide type errors. This is the same tradeoff as any * node - anything connects to anything, so a CONDITIONING accidentally wired where a LATENT belongs won't be caught until it isn't. Wire carefully; the graph won't save you.

Install

Search "Model Phase Sync" in ComfyUI Manager, or:

comfy node install model-phase-sync

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/josetseph/ComfyUI-ModelPhaseSync.git

Restart ComfyUI. The node is Sync Barrier, under the model_phase_sync category. No dependencies, no model downloads, MIT licensed. The pack ships six example workflows (Z-Image barrier batching and the three-stage CLIP→diffusion→VAE split) if you'd rather start from a working graph than hand-wire one.