ComfyUI v0.43.0: Kling 4.0 Video Nodes With Stereo Audio

• Kling 4.0: text to video, image to video, reference to video and edit video nodes, with clips up to 30 seconds and stereo audio.
• Meshy T2.5: Smart Topology for text to 3D and image to 3D generation.
• PixVerse V6 Max: a new model, with updated V6 price badges.
• Fibo 2.0: text to image and image editing support for the open weights.
• MOSS-Image: LoRA loaders apply to the edit model, with attention bias, image strength and rope scaling controls.
• New nodes: Branch, Pad Image and a MOSS model merge node.

Kling 4.0 Nodes Generate Clips Up to 30 Seconds With Stereo Audio #

Kling 4.0 adds four node types: text to video, image to video, reference to video and edit video. Each clip can run up to 30 seconds, and stereo audio is generated in the same pass as the video, as described in the ComfyUI v0.43.0 changelog. Because the audio arrives with the clip, this suggests that a graph which added a separate audio step after Kling output would be affected by this change. The entries do not say how such a step behaves alongside the generated audio.

Meshy T2.5 Adds Smart Topology to 3D Generation; PixVerse V6 Max Is a New Model #

Meshy T2.5 adds Smart Topology to both text to 3D and image to 3D generation, so 3D graphs built on either input get the option. PixVerse V6 Max is a new model, and the V6 price badges were updated in the same release. Both entries appear in the ComfyUI v0.43.0 changelog with no parameter details, so graphs that use these nodes need their output checked by the reader.

Fibo 2.0 Open Weights Add Text to Image and Image Editing #

Fibo 2.0 gains text to image and image editing support for its open weights, so those two tasks can now run on the open weights inside ComfyUI. The changelog does not state how the open weights compare with partner model calls, so the results and any change to existing graphs are not established by the release notes. The entries also do not list the hardware the open Fibo 2.0 weights need, which a reader must establish before running them locally.

MOSS-Image Edit LoRAs and Edit Controls #

LoRA loaders now apply to the MOSS-Image edit model, so edit LoRAs can be attached through those loaders. Attention bias adds conditioning for the edit model, and image strength adds a control for how strongly the source image carries over into the edit. Rope scaling adds support for larger MOSS-Image resolutions, as listed in the ComfyUI v0.43.0 changelog.

Branch Node Selects From Connected Inputs by Name #

The Branch node selects from connected inputs by name, and its inputs can be renamed. The ComfyUI v0.43.0 changelog describes it in one line:

"Select from connected inputs by name, with inputs you can rename"

ComfyUI v0.43.0 changelog

Renamed inputs give each connected path a name the reader can read in the graph. The entry does not describe how the node chooses between inputs when several are connected, so that behaviour should be confirmed in a graph before it is relied on.

Pad Image Adds Solid Color, Mirrored Edges or Padding to a Fixed Multiple #

Pad Image pads an image with a solid color or a mirrored copy of its edges, and it includes padding to a fixed multiple. Mirrored padding fills the border with a copy of the image's edges, while a solid color fills it with one color. The changelog lists the fixed-multiple option without describing which models need it or what sizes it produces.

Bug Fixes That Change Existing Graph Behaviour #

Two Fibo 2.0 fixes change how existing text to image graphs behave: text to image negative prompts are now handled correctly, and RoPE scaling is now NTK aware. Gemma 4 31B global attention now uses k_eq_v with 4 global KV heads, and the top-p sampling threshold is now applied where it should be. Graphs that depend on any of these settings are the ones whose results can change.

Offload streams are now synced before a pin is stolen, which affects graphs that run offloaded models. Pytorch attention is no longer used for head dims above 256 on gfx11, and MOSS now has a lower peak memory when it runs. The gfx11 change applies to users of those GPUs, and the MOSS change applies to graphs that run MOSS.

Partner Nodes Behind a Proxy Now Respect Proxy Environment Variables #

The partner node client now respects proxy environment variables, so partner node calls made from behind a proxy follow the proxy settings they already define. Users who reach partner nodes through a proxy are affected by this change. The entry describes the fix only, so the reader should confirm partner node calls after upgrading.

Also in This Release #

• Bria guided nodes: Guided Remove Background, Product Holding and Virtual Try-On.
• Ideogram Vector Generate SVG: SVG vectors from a prompt, or an image traced into one.
• MOSS model merge: a model specific merge node for MOSS.
• ACEScg color space: Convert Image Color Space and Save Image (Advanced) accept ACEScg for EXR output.
• Qwen 3.x performance, part III: adaptive tree MTP.
• Comfy-Aimdo 0.6.0: dynamic VRAM KV caches, part I.
• Flux2 VAE: fused and stripped through comfy-kitchen.
• Pytorch attention in the VAE on AMD.
• Templates v0.11.96 and v0.11.97, Embedded docs v0.5.15, Frontend 1.57.4.

The ComfyUI v0.43.0 changelog has the full entries for each item above.