In this article
ComfyUI v0.46.0 removes DirectML support
ComfyUI v0.46.0 removes DirectML support, so check your setup before you upgrade if you run on it. The release also adds partner and model nodes and new image and video tools, and its changelog compares it against v0.45.0.
New nodes, video tools and model support
Partner and model node changes:
- LTX: new AlphaGen, SDR to HDR, Reframe and Upscale nodes.
- Kling 4.0: Image and Long Image nodes added, plus 4:3 and 3:4 aspect ratios.
- HeyGen: Video 1.0 nodes gain 1080p output, 30 second clips and automatic duration, and Avatar 6 nodes are added.
- Language models: the Anthropic Claude node adds Claude Fable 5.2, and the Gemini-LLM node adds Hydrogen, Argon and Carbon.
- Other additions: a MiniMax H3 Image node, RMBG 3.0 in the Bria background-removal node, and Seedance 2.5 Flash.
Core and local tooling also changed:
- Video: a new Video Deflicker node and a Faster RAFT optical flow implementation.
- Images: Save Image advanced can save single channel PNG files, and a new node splits image channels.
- Qwen: embedded tokenizers in vocabulary-pruned checkpoints are supported, and Qwen3-1.7B works in the Generate Text node.
- Other support: PAN2 is supported, and SenseNova gains thinking, interleaved generation and self-correction.
- Concatenate Video: clips now conform to the first clip's frame rate and audio.
Removed models and rerouted partner nodes
Claude Sonnet 4.5 is gone from the partner nodes, along with leftover Moonvalley code. The OpenRouter node dropped the retired Qwen3.6 Max Preview model and added Meta Muse Wonder, Spark 1.3 and Ember. Wan, Krea, Luma and more partner nodes moved to Comfy Router, and FishAudio speech-to-text now uses transcribe-1-pro.
A CUDA 13 warning, a new flag and fixes
ComfyUI now shows a yellow warning when a GPU that supports CUDA 13 runs without it. The new --force-output-rerun flag makes output nodes write their files again on a cache hit. The commands and flags here come from the release notes listed under Sources, so confirm them against your own install before running them.
Kling Avatar and Topaz nodes now show task or request IDs when polling times out. Fixes include a dynamic VRAM buffer restore that no longer leaks a sampling schedule between models, and a corrected APG output at cfg 1.0 in some cases.
Google Omni now downloads the generated video even when Google marks the file as failed. The bundled frontend package is updated through 1.59.6, workflow templates through v0.11.111 and embedded docs to v0.5.17.
What the ComfyUI notes leave out
The release notes give no upgrade or migration steps for users who relied on DirectML. They also give no minimum CUDA or driver version for the new CUDA 13 warning. They list no pricing or rate limits for the new partner and model nodes. They do not say which model or hardware changes affect memory requirements, and the release page gives no release date.
Who should check what before upgrading
If you run ComfyUI on DirectML, hold off on upgrading until you have another backend to move to. If your workflows use models that were removed from the partner or OpenRouter nodes, check them before you upgrade.
Sources
Image: Photo by Bhautik Patel on Unsplash
This story was produced with AI assistance from the sources listed above and fact-checked against them before publication. Read our editorial methodology and the corrections log. Spotted an error? Email [email protected].





Discussion