ComfyUI Kitchen vs SageAttention: setup and video tests

This comparison tests Comfy Kitchen attention against SageAttention in ComfyUI using MiniMax H3 text-to-video workflows. The presenter describes Kitchen as an INT8 attention kernel from Comfy Org's GPU kernel library, included as a normal ComfyUI dependency, with Windows and AMD support.

The setup explanation covers selecting Kitchen through the native model attention backend node or enabling it globally at startup. The presenter recommends the node for control over individual models and warns that quantized attention can change output, so users should test their own workflows.

For a 10-second video at 20 sampling steps, the reported times are 2 minutes 54 seconds for Kitchen and 2 minutes 48 seconds for SageAttention. At 30 steps without the turbo LoRA, the reported times are 4 minutes 31 seconds and 4 minutes 24 seconds respectively. These are results from the presenter's setup; the transcript does not specify the GPU, and the speaker cautions that other hardware will produce different timings.

The presenter prefers Kitchen's detail in parts of the action and cooking examples while describing motion as similar. An FPS mismatch in one comparison is corrected to 24 before synchronizing the clips. The shared workflow lets viewers compare attention modes and toggle the turbo LoRA for their own local AI video tests.