ComfyUI v0.43.0 Adds MOSS-Image and Fibo 2.0 Support
The latest ComfyUI update introduces expanded model support, refined memory management, and several AMD-specific performance improvements alongside a host of new integration nodes.
ComfyUI v0.43.0 shipped on October 27 with MOSS-Image and Fibo 2.0 support, alongside memory-handling changes and AMD-specific fixes covering changes since v0.42.0.
MOSS-Image additions include rope scaling, color-depth control, edit LoRA support, attention bias conditioning with image strength, a model-specific merge node, and lower peak memory use. Fibo 2.0 arrives as a core feature with NTK-aware RoPE scaling and a negative prompt fix. MiniMax-H3 PDD head banks get single-head LoRA support, while Gemma 4 31B global attention is fixed. Flux2 VAE can now be fused and stripped via comfy-kitchen.
Comfy-Aimdo bumps to 0.6.0. Dynamic VRAM KV caches part one and Qwen 3.x adaptive tree MTP part III land. AMD hardware enables PyTorch attention in the VAE, excluding head dimensions above 256 on gfx11.
Partner API nodes now renew sign-in tokens and the client respects proxy environment variables. Integrations span HeyGen, ClarityAI, Black Forest Labs, Bria, Krea, VEED, Mirelo, Meshy, PixVerse, Ideogram, and Kling. Core additions introduce a Pad Image node, a Branch node, ACEScg support for EXR nodes, deploy-to-cloud endpoints behind a feature flag, and Comfy API console tips.
The optional local agent loads through comfy-agent's v1 host. The frontend package moves to 1.57.4, workflow templates to v0.11.97, and embedded docs to v0.5.15. Four first-time contributors joined.