Skip to content
AIBRAINWORX / OpenSourceCoder.comLOCAL-FIRST. BUILT IN THE OPEN.
Open Source / Developer & Agent Tools

ComfyUI 0.46.0 fixes VRAM-state leakage and clarifies the cost of repeated outputs

The release combines local execution fixes with hosted partner nodes. An opt-in output rerun flag and removal of DirectML support deserve attention before an upgrade.

Listen to this articleBrowser Voice
Ready
The ComfyUI organization’s official project icon.
Official ComfyUI project-owner GitHub avatar, used for product identification; no endorsement or relicensing of the mark.
Event / context: 2026-11-18 UTC. Source check: Nov 20, 2026.

ComfyUI 0.46.0 brings several changes that matter to repeatable image, video and audio workflows: a dynamic-VRAM state-restoration fix, an opt-in way to rerun output nodes after cache hits and removal of DirectML support. The release was published on November 18 UTC. We reviewed the upstream changes without updating or running a local ComfyUI installation.

The release notes also list new and updated partner nodes. Those additions should be read separately from local model support: an interface to a hosted service does not make that provider’s models downloadable or its generation free.

Keeping one model’s state out of another run

The dynamic-VRAM fix addresses model-buffer restoration that could leak a sampling schedule between models. The contributor supplies regression examples and reported tests, but we have not reproduced them. This is a specific state-management correction, not a claim that every previous inconsistent image came from the same bug.

Our analysis: graph-based workflows are often compared by their visible nodes and seeds, but execution history can matter too. If a backend restores the wrong internal state, two apparently equivalent requests may not be equivalent in practice. Keeping the application version and loaded-model sequence in a test record can help distinguish that problem from ordinary sampling variation.

Cached computation and missing output files

A new opt-in force-output-rerun flag tackles another mismatch: a repeated request can reuse cached execution after an external process has deleted the saved output file. The change lets output nodes write their files again while keeping upstream computation cached.

The implementation description explicitly excludes API nodes, because rerunning one could repeat a paid external request. That boundary is important for anyone hosting a workflow service. A “successful” cached result is not useful if its file is gone, but blindly repeating every node can introduce a different cost problem. The flag is off by default; it is not a new default behavior for all local users.

Check the backend before upgrading

The DirectML removal explains that the supported PyTorch versions have moved beyond the last version with that plugin. Users relying on that execution path should verify an alternative before replacing a working environment. This does not mean all AMD hardware or all Windows installations lose support; backend and device compatibility are separate questions.

ComfyUI remains an open-source project under its GPL license. Individual model weights, custom nodes and hosted services retain their own terms. Our ComfyUI repository profile links to the exact upstream project so readers can inspect the implementation and release rather than infer compatibility from the version number alone.

EXPLORE THE SOURCE

GitHub repositories

Original reporting from primary source review. Publisher evaluations are attributed; no installation, local inference, device test or independent benchmark was performed.

Follow your curiosity.

Pass the signal.