Stop ComfyUI from reloading CLIP / UNet / VAE between branches when you batch multiple prompts in one workflow. Force true phase barriers so each model loads once, does all of its work, then yields to the next phase.

The problem

Batched multi-prompt graphs (several CLIP encodes → several samplers → several VAE decodes) often thrash on constrained VRAM / unified memory (especially Apple Silicon): CLIP loads, gets partially unloaded for UNet, reloads, unloads again… minutes of waste per run.

The fix

| Node | What it does | |------|----------------| | Sync Barrier | True execution barrier. Outputs only resolve once all connected inputs are ready, so Comfy finishes an entire phase before the next starts. Sockets auto-grow as you wire branches (up to 1000). | | Phase Unload | Passthrough that calls unload_all_models() after a barrier. Required on Apple Silicon: Comfy’s auto-evict treats macOS “available” RAM as free VRAM, so CLIP/UNet otherwise stay resident into the next phase. | | Save / Load Conditioning | Checkpoint CLIP CONDITIONING to disk (and reload it) so you can run text encode as its own workflow, then sample later without CLIP in memory. Writes are atomic (temp → os.replace) so watchers never see a half-written file. | | Load Latent (Upload) | Same idea for .latent files — built-in Load Latent has no upload button; this one does. Compatible with ComfyUI Save Latent. | | Save Latent / Save Image (Atomic) | Drop-in style saves that finish with os.replace, for processes that harvest files as soon as the path appears. |

Before / after (3-image batch, Z-Image Turbo on MPS)

| Approach | Wall time | Behavior | |----------|-----------|----------| | 3 separate single-image runs | ~1500s | One clean load cycle each, no batching | | Batched, no barrier | ~1286s | Repeated Unloaded partially / reload thrash | | Batched + Sync Barrier | 887s (14:47) | One CLIP load, one UNet load, one VAE load |

Install

ComfyUI-Manager / Registry

Search for Model Phase Sync (or model-phase-sync) on the Comfy Registry / Manager:

comfy node install model-phase-sync

Manual

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

Restart ComfyUI. Nodes appear under model_phase_sync.

How to use

Duplicate your per-prompt branch (CLIP encode, sampler, etc.) for each prompt.
After each phase’s branch outputs (e.g. all CLIP CONDITIONINGs), insert Sync Barrier.
Wire in_1…in_N from each branch; pull out_1…out_N into the matching next-phase nodes.
New empty sockets appear automatically as you connect — you never scroll past hundreds of unused pins.
After each barrier, put Phase Unload on every wire into the next phase (each out_i → Unload → sampler/VAE). One unload on a single branch does not gate the others. Do not hang unload off an individual save: with many saves it can fire mid-phase.

Split phases across workflows

You do not have to keep CLIP, UNet, and VAE in one graph. Checkpoint between phases:

1) CLIP only
   CLIPTextEncode → Save Conditioning  →  output/conditionings/

2) Diffusion only
   Load Conditioning → KSampler → Save Latent  →  output/latents/

3) VAE only
   Load Latent (Upload) → VAE Decode → Save Image

Save Conditioning writes a .conditioning file (passthrough so you can still wire a barrier in the same graph). Load Conditioning and Load Latent (Upload) each have a choose file to upload button (Comfy’s built-in Load Image upload only works for images; stock Load Latent has none). Uploads go to input/conditionings/ or input/latents/. You can also use the dropdown or paste a path. Use distinct filename_prefix values per prompt (e.g. conditionings/prompt_01, latents/zimage_01).

That way each workflow only uses the model it needs — but Comfy keeps models loaded across /prompt calls. On MPS it often will not evict the previous phase (CLIP can sit through all of diffusion). Do not put an unload node in these graphs: with many saves it can run halfway through the batch, and at the start of the next graph it can race the next loader.

After each phase prompt finishes (history/success, not immediately after you POST /prompt), free memory:

POST /free
{"unload_models": true, "free_memory": true}

Call that between CLIP → diffusion and diffusion → VAE. In the GUI, that is the Free memory button. Queuing the next workflow alone is not enough. Barriers still help when you batch many branches inside one workflow.

Conditioning save/load is inspired by the idea behind ComfyUI-SaveAndLoadPromptCondition (not on the Registry); this pack ships a clean MIT implementation aimed at phase-split batching.

Example workflows

One-shot / barrier batching

| File | Role | |------|------| | workflows/Z-Image-Sync-Barrier-Example.json | Full pipeline in one graph: CLIP barrier → Phase Unload → sample → UNet barrier → Phase Unload → VAE → Save Image | | workflows/Z-Image-Sample-And-Checkpoint-Latents.json | Sample + checkpoint: CLIP barrier → Phase Unload → KSampler → Save Latent per branch (no VAE in this graph) | | workflows/Z-Image-Decode-Checkpointed-Latents.json | Finalize: Load Latent → VAE Decode → Save Image (no samplers). After the sample workflow finishes, POST /free before this one. |

Split phase pipelines (CLIP → diffusion → VAE)

Run these in order. POST /free (or Free memory) after each stage completes — these graphs have no unload node on purpose.

| File | Role | |------|------| | workflows/Z-Image-Phase-CLIP-Save-Conditioning.json | CLIP: encode prompts → Save Conditioning | | workflows/Z-Image-Phase-Diffusion-Save-Latents.json | Diffusion: Load Conditioning → KSampler → Save Latent | | workflows/Z-Image-Phase-VAE-Save-Images.json | VAE: Load Latent (Upload) → VAE Decode → Save Image |

Why latent checkpoints exist (inside one sample workflow)

A UNet Sync Barrier waits until every sampler finishes before any VAE/Save Image can run. That is what stops model thrash — and it also means no PNGs on disk until the whole sample phase is done. If ComfyUI crashes mid-batch, finished branches only existed as in-memory latents.

KSampler_i ──► Save Latent_i ──► UNet Barrier ──► (later) VAE / Save Image

Each Save Latent depends only on its own sampler, so it writes as soon as that branch finishes. Prefer distinct prefixes (e.g. latents/zimage_01). Decode later with the finalize / VAE-phase workflow — no resampling.

Publishing / updates

This pack publishes to the Comfy Registry via GitHub Actions whenever version in pyproject.toml changes on main. Maintainers: set the repo secret REGISTRY_ACCESS_TOKEN to your Registry API key (docs).

License

MIT — free to use, modify, and redistribute. Please keep the copyright notice and credit josetseph.

GitHub repo tips (maintainers)

Description: Stop ComfyUI from reloading CLIP/UNet/VAE between branches when batching multiple prompts in one workflow.
Topics: comfyui, comfyui-custom-nodes, model-management, vram-optimization