Skip to content

ComfyUI Deleted the DirectML Flag a Year After Warning You Off It

ComfyUI 0.46 removes --directml and the whole torch-directml path, 69 lines across five files. The reason is a torch version that fell out of support, not a change of heart. What breaks on upgrade, what to pin, and the second change in the same window that nobody put in the headline.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
An open desktop computer case on a workbench with a graphics card resting outside it, an empty slot visible inside, warm daylight and dust
Editorial illustration; not a screenshot of the software.

Version 0.46.0 of ComfyUI shipped on November 18. In the changelog, between a Deflicker node and a batch of partner-node additions, there is a line that reads Remove directml support. Five files changed, 69 lines deleted, 9 added. That is the entire visible footprint of the change.

The interesting part is not the deletion. It is that the project had been announcing it in a log message for over a year.

Anyone who launched with --directml got this at startup, reprinted from comfy/model_management.py:

WARNING: torch-directml barely works, is very slow, has not been updated in over 1 year and might be removed soon, please don't use it, there are better options.

That warning is now deleted along with the flag, the directml_enabled branch that six other call sites checked, and the latent-preview special case that existed only because DirectML could not hold a tensor type. A tool that tells you it is broken for a year and then removes itself is behaving reasonably. It is still a removal, and removals land on people who were not reading.

What DirectML was for

DirectML was the path that let ComfyUI run torch on Windows without CUDA. In practice that meant AMD and Intel GPUs on Windows machines, which is a large group of people who bought a graphics card for reasons other than stable diffusion and later discovered they wanted to generate images with it.

It was always the slow lane. The warning said so at every launch, in the tone of someone who had given up being polite about it. But it worked, and for a while it was the only thing that worked without changing operating system, buying different hardware, or renting a GPU by the hour.

The reason for removal is narrower than the warning suggests. The pull request body, dated November 12, says: "The last pytorch version that had a directml plugin was pytorch 2.5 and we don't support that anymore." Nothing about DirectML changed. The torch version that carried it left the support window, and maintaining a backend on an unsupported torch build is a cost with no upside.

What happens if you upgrade without looking

If your launch script passes --directml, 0.46 rejects the argument. The error looks like a typo, which is the worst kind, because the fix is not a typo but a migration decision.

Three things disappear with it: the CLI flag, the device-selection branch in model_management.py, and the mention of DirectML in the accelerator list that API nodes advertise. If you had a wrapper script, a systemd unit, or a .bat file from two years ago, check it before you pull. I have been bitten by exactly this pattern with other tools: the flag lives in a file you have not opened since you wrote it.

Rolling back is technically possible and practically a dead end. Pin ComfyUI to 0.45.x, keep torch 2.5, keep torch-directml. That stack is unsupported twice over: 0.45.0's own README says torch 2.7 is the minimum it supports, and torch-directml never shipped for anything newer than 2.5. So the rollback is a combination no upstream project tests, and it freezes on November 10, 2026. No model support added after that date, no VRAM fixes, no new nodes, no partner nodes. For a machine that runs one fixed workflow to produce one fixed kind of asset, a frozen install is a legitimate engineering choice. For anything current, you are now maintaining a fork without calling it one.

If you need those GPUs, the list is short and I got one entry wrong the first time. Intel cards still have a path, since --oneapi-device-selector survives in cli_args.py. AMD cards on Linux with supported silicon can use ROCm. Windows also has a documented ROCm path now, which I had written off: the 0.46.0 README walks through AMD's multi-architecture PyTorch wheels, and the requirements are Windows 11, a current AMD driver, 64-bit Python 3.13, and an RDNA 2, RDNA 3, RDNA 3.5 or RDNA 4 card. The README's example pins torch 2.13 with ROCm 10.0, so this is a different PyTorch build and a fresh install rather than a flag you flip on the old one. It is a path, not a drop-in, and the difference matters if you were about to tell someone their card is finished.

The other change in the same window

Two weeks before the removal, on November 5, 0.45.0 turned local_web_sso and partner_node_token on by default. The pull request that did it explains why it needed a flag before: the flag registry had a default for every feature flag, but only flags passed with --feature-flag reached the server, so nothing could actually be on by default. Until that landed, a self-hosted user had to start with --feature-flag local_web_sso --feature-flag partner_node_token to get sign-in at all.

Now the default is on: the frontend offers browser sign-in through a Cloud popup. The flag's own description scopes that to a Local ComfyUI opened at localhost or 127.0.0.1, so a server reached by its LAN address is not the case this covers. If your local install sits behind a firewall, on a network with a proxy that eats outbound requests, or on a machine you deliberately keep offline, it is still the change to test first. The failure mode is not a crash. It is a sign-in that quietly never completes.

The same release window also moved several partner-node families to something the notes call Comfy Router. Qwen Image went first, in 0.45.0. Wan, Krea and Luma followed in 0.46.0, along with a catch-all pull request whose title is simply "move more partner nodes to Comfy Router". The release notes name the router and never define it, so I will only state what is visible: the node families that talk to outside providers are being consolidated behind Comfy's own routing layer instead of each family reaching its provider directly. Whether that changes your keys, your egress or your billing is something to verify on your own install rather than infer from a changelog line. What is not in doubt is the direction. Fewer direct provider paths, more brokered ones, and account plumbing switched on by default.

Three small lines worth more than their length

The CUDA 13 warning. 0.46 adds a warning when your GPU supports CUDA 13 but your torch build is not using it. It was softened to yellow in a follow-up commit the same week, which is a reasonable call, because it is not an error. But a yellow line you learn to scroll past is worse than a red one you fix. If you see it, your install runs, and it runs below what the card can do.

A VRAM buffer leak between models. The fix lands in 0.46 for a bug where restoring the dynamic VRAM buffer leaked a sampling schedule between models. If your second generation in a batch has ever come out slightly different from the first with the same prompt and seed, this is the class of bug that explains it. It is the kind of fix nobody notices, which is the point.

Deflicker and faster optical flow. A video Deflicker node and a faster RAFT implementation, both from kijai, land in the same release. For anyone stitching a sequence out of separately generated clips, a deflicker pass is the difference between a cut and a strobe. This one is a feature, not a removal, and it is the reason some people will upgrade on purpose.

An upgrade pass that takes ten minutes

  1. Write down what you actually have, not what the folder is called. git -C ComfyUI describe --tags for the version, and python -c "import torch; print(torch.__version__, torch.version.cuda)" for the stack underneath. On the Windows portable, that python is python_embeded\python.exe inside the install folder, not whatever a PATH python resolves to.
  2. Grep your launch scripts for flags that no longer exist. Start with --directml, then any --feature-flag you added by hand, then anything you copied from a forum post in 2024.
  3. Produce one baseline output before touching anything. Fixed workflow, fixed seed, saved file. Without it, step 4 is vibes.
  4. Upgrade, rerun the same workflow with the same seed, and compare. Not a glance at the thumbnail: open both at full size and look at the detail that usually fails on your model.
  5. Read the removals instead of the features. Remove directml support was one line in a list of forty partner-node additions, and it affects more people than all forty.
  6. Choose pin or migrate on purpose. Pinning keeps a working job working. Migrating costs a weekend and gives you a machine that still gets fixes in six months.

None of this is an emergency. 0.46 is a normal release, and most of it is additive. What makes it worth a careful read is that three separate edges of the tool moved in a fortnight: a backend went away, sign-in offered itself by default on loopback installs, and provider calls started consolidating behind a router. Two of those three get no paragraph in the announcement. If you maintain a local install for anyone besides yourself, that is the release to read line by line.

Sources

Share this guide
X LinkedIn