Introduction

ComfyUI v0.43.0 is out. The GitHub Release went live at 02:40 UTC on October 27 (10:40 MYT), eight days after v0.42.1, and it packs 57 commits into one build you can pin.

If you've been following our ComfyUI coverage, a lot of the headline names will look familiar. Kling 4.0, FLUX 3 Max, Fibo 2.0 and MOSS-Image editing all landed on the master branch over the past week. v0.43.0 is the first tagged release that carries them, so Windows portable users and anyone pinned to tags finally get them too. Along the way it also slips in some quieter upgrades that make everyday graphs nicer to build.

Release: ComfyUI v0.43.0 · Compare: v0.42.1…v0.43.0

Grab the build

Five Windows portables are attached to the tag:

  • ComfyUI_windows_portable_nvidia.7z
  • ComfyUI_windows_portable_nvidia_cu126.7z
  • ComfyUI_windows_portable_nvidia_arm64.7z
  • ComfyUI_windows_portable_amd.7z
  • ComfyUI_windows_portable_intel.7z

Git installs can pin the tag directly:

cd ComfyUI
git fetch --tags
git checkout v0.43.0
pip install -r requirements.txt

The requirements bump pulls in workflow templates v0.11.97, so the template browser picks up the newest example graphs too.

Partner nodes: a big week

Partner nodes are the paid API nodes billed on Comfy credits. This release adds a lot of them:

  • Kling 4.0 (#17364): Text to Video, Image to Video with up to 10 keyframes, Reference to Video, and Edit Video, all with audio, 720p or 1080p, 3 to 30 seconds. Full breakdown in Kling 4.0 Lands in ComfyUI Partner Nodes.
  • FLUX 3 Max (#17202), plus 2K and 4K output on the FLUX 3 Video nodes (#17200). See FLUX 3 Max in ComfyUI.
  • Krea 3 Image gains an ultra thinking mode toggle (#17260).
  • PixVerse V6 Max (#17287), with updated V6 price badges.
  • Ideogram Vector Generate SVG and Image to SVG nodes (#17298).
  • Mirelo SFX Video to SFX and Text to SFX (#17265).
  • Bria FIBO 2.0 Image Generate, with Image Edit moved to FIBO 2.0 (#17229), plus three new Bria nodes: Guided Remove Background, Product Holding and Virtual Try-On (#17289).
  • VEED Fabric 1.0, Lipsync 2.0, Background Removal and Clean Audio (#17242).
  • HeyGen Voice 1.0 instant voice clone and text-to-speech (#17043).
  • ClarityAI Crystal Video Upscaler (#17192).
  • Meshy T2.5 Smart Topology nodes for cleaner 3D meshes (#17280).

The partner-node client now respects proxy environment variables too (#17346), which helps if you run ComfyUI behind a corporate or studio proxy.

Local models: Fibo 2.0 and MOSS-Image edit

Not everything here needs an API key.

  • Fibo 2.0 runs natively (#17240, by kijai), with follow-ups for NTK-aware RoPE scaling (#17290) and correct negative-prompt handling in text-to-image (#17285). Background on the model is in Bria Fibo 2.0.
  • MOSS-Image keeps growing: edit LoRAs load (#17296), edit mode gets attention-bias conditioning and an image strength control (#17320), color-depth control arrives (#17206), RoPE scaling lands (#17172), and there is a model-specific merge node (#17349). Peak memory also drops a little (#17338). See MOSS-Image Edit LoRAs in ComfyUI.
  • MiniMax H3: single-head LoRAs now apply to the PDD head banks (#17239).

New core nodes for graph builders

These are small, but you'll reach for them constantly.

Branch node

The new Branch node (#16383) takes any number of inputs and lets you pick which path runs by name. It's a clean way to keep "draft" and "final" routes, or several style variants, in one graph and flip between them without rewiring.

Pad Image with mirror padding

Pad Image (PadImageNode, #17224) pads by fixed dimensions, by a multiplier, or up to a multiple. That last mode exists because models like LTX want width and height in multiples of 32. You can pad with a solid color (with feathering) or mirror the image. The PR notes that solid-color padding tends to turn into hard edges in the output, while mirror padding blends much better. Pad, generate, then crop back to your exact resolution.

ACEScg for VFX pipelines

Convert Image Color Space and Save Image (Advanced) now support HDR ACEScg (#17023). Before this, the only EXR route was linear Rec.709, which clipped any color outside that gamut. Now you can save scene-linear ACEScg EXRs tagged with AP1 primaries and D60 white, the standard handoff format for ACES compositing.

Speed and memory

  • AMD VAE attention (#17271): the VAE now uses PyTorch attention on AMD GPUs where AOTriton works. On an RX 9060 XT, a Wan 2.1 VAE decode at 2048² dropped from 8.28 GiB peak to 1.64 GiB, and encode from 8.28 GiB to 1.41 GiB. Decode time went from 2.02 s to 2.44 s, a small price for that much headroom.
  • Flux2 VAE fused and stripped through comfy-kitchen (#17196): it now runs in fp16, which the PR measures at about 77 dB decode PSNR against fp32, compared with about 59 dB for the old bf16 default. Peak VRAM is also lower and it runs faster.
  • Comfy-Aimdo 0.6.0 and Dynamic VRAM KV caches, part I (#17274): KV caches for in-graph LLMs can now spill between VRAM and RAM, and Windows MCDM support is included.
  • Qwen 3.x performance, part III (#17273) adds adaptive tree MTP for faster local prompt enhancers.

Deploy to Cloud (behind a flag)

#17203 adds the plumbing for Deploy to Cloud. A local ComfyUI can bundle its setup (snapshot, model list, and an optional workflow), send it to Comfy's builder, and get back a link that opens the Create a Build wizard already filled in. It ships behind a feature flag, it only sends anything when you explicitly ask, and it stays off entirely under --offline. This is the same deployment path apps like OPENjutsu lean on.

Pro Tip

If you run AMD, update for the VAE fix alone. If you build big graphs, drop a Branch node in front of your sampler so draft and final settings live side by side.

Original Source

Conclusion

ComfyUI v0.43.0 is a stacked release. It moves Kling 4.0, FLUX 3 Max, Fibo 2.0 and MOSS-Image edit from master into a tag you can pin, adds about a dozen new partner nodes, and gives graph builders a Branch node, mirror padding and real ACEScg output. AMD users get a VAE that needs roughly a fifth of the memory. Update, open the template browser, and go play.

Want to try FLUX right now without installing anything? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=comfyui-v0-43-0-release

—Aurelia ♡