ComfyUI v0.32.0: New Native Attention Backend and LTX 2.5 Support

ComfyUI Wikinews

ComfyUI v0.32.0 adds a built-in attention backend matching SageAttention speed, native LTX 2.5 support with STG and dual CFG, MiniMax H3 VAE and memory fixes, plus partner nodes.

ComfyUI v0.32.0 is now available. The release adds a new built-in attention backend (comfy kitchen attention) that reaches SageAttention-class speed without any extra installs, native LTX 2.5 support, MiniMax H3 VAE and memory optimizations, and updated partner nodes including Qwen Image 3.0 and Grok Imagine 2.0.

New Built-in Attention Backend

v0.32.0 implements comfy kitchen attention, a native attention backend that ships with ComfyUI itself. A new ModelAttentionBackend node lets you select the attention implementation per model directly inside a workflow, currently offering PyTorch attention or comfy kitchen attention. For users who want it everywhere, the --use-ck-attention startup argument enables it as the default backend for all models (with the caveat that it might break some).

Early community benchmarks put it on par with SageAttention. On MiniMax H3 it matched SageAttention Auto, and in a Z-Image Turbo 2048x2048 @ 9 steps test the new backend took 14.55s, versus 14.24s for SageAttention (Auto) and 23.16s for the default PyTorch attention:

Attention backendZ-Image Turbo 2048x2048 @ 9 steps (3-run average)
Comfy Kitchen Attention14.55s
SageAttention (Auto)14.24s
PyTorch Attention (default)23.16s

For users who have trouble installing SageAttention, this removes that friction entirely since the backend is built into ComfyUI.

Native LTX 2.5 Support

LTX 2.5 is now natively supported in ComfyUI with spatio-temporal guidance (STG), dual CFG, and duration prediction. This release also adds the new LTXV nodes:

  • LTXV Spatio-Temporal Guidance: STG guidance for LTX video generation
  • LTXV Modality Guidance: audio/video coupling guidance for LTX
  • LTXV Dual CFG Guider: dual-CFG guider for LTX workflows
  • LTXV Duration Predictor: predicts a natural shot duration from caption tokens

MiniMax H3 Optimizations

Several changes in v0.32.0 target the MiniMax H3 ecosystem:

  • Optimized MiniMax-H3 VAE for faster encoding and decoding
  • Fixed peak memory issue with MiniMax H3 during generation
  • Fixed VAEDecodeTiled crash on NestedTensor latents (MiniMax H3)

Partner Node Updates

  • Qwen Image 3.0: added Qwen-Image 3.0 and 3.0 Pro text-to-image and edit nodes
  • LTX 2.5: added LTX 2.5 Text/Image/Audio to Video partner nodes
  • Grok Imagine Image 2.0: added grok-imagine-image-2.0 model support

Performance and Stability

  • PyTorch 2.7 is now the minimum officially supported PyTorch version
  • ER-SDE noise scaler extended by scaling h(t)
  • Mistral and Llama tokenizers no longer depend on transformers
  • Fixed upscale models breaking on non-dynamic low VRAM
  • Fixed broken tiled audio decode
  • Fixed CLIP Vision regression
  • Made Create Layered Image discoverable with clearer flags

For the full changelog, visit the official GitHub release or the ComfyUI docs changelog.

Comments

Sign in with GitHub to join the discussion.

Loading comments…