ComfyUI-AppleSilicon-int8-Fast

This custom node is based on the original work of Paweł Mazurkiewicz (ComfyUI-AppleSilicon-FP8).

A lot of speed-up ideas were taken from the amazing work of Liu Liu on Draw Things. Open source repos:

• draw-things-community (Draw Things CLI)
• ccv
• s4nnc

Proudly vibe-coded using Claude.

ComfyUI-AppleSilicon-int8-Fast

Fast int8 support in ComfyUI on Apple Silicon.

Copy this folder into ComfyUI/custom_nodes/ and restart ComfyUI.

Alternative: Open the terminal, cd to your custom nodes folder, type git clone https://github.com/Illuminus7/ComfyUI-AppleSilicon-int8-Fast

The kernels are tuned on an M5 Max. Smaller M5 chips run the same kernels with fewer GPU cores; they should get the same relative speedups, but the large models (MiniMax H3 needs about 40 GB) won't fit in their memory.

This was tested with ComfyUI v0.43 and PyTorch 2.13 to 2.15. Future updates may break compatibility.

No startup arguments are needed (Comfy Desktop's default --enable-manager is fine). Memory-related arguments are not recommended.

The node pack adds one node, Load LoRA (Bypass, Optional): ComfyUI's bypass LoRA loader with None as the first choice, so a workflow can keep a LoRA slot and still run without a LoRA.

Sample workflows for Krea 2, MiniMax H3 and LTX 2.5 are on Hugging Face: Illuminus7/Comfy_Int8_Fast_Workflows. Using the exact models in the workflows is recommended. Put each file in the listed folder under ComfyUI/models/. The image-to-video workflows need a start image of your own in the Load Image node.

INT8 attention is on by default. It makes generation faster (about 6% for Krea 2 at 2048×2048 and 9–12% for MiniMax H3 video), but it may change images slightly: same scene and composition, with small differences in details. To turn it off, start ComfyUI with the environment variable INT8_ATTENTION=off:

• from the terminal: INT8_ATTENTION=off python main.py
• ComfyUI Desktop: run launchctl setenv INT8_ATTENTION off in the terminal, then restart ComfyUI (launchctl unsetenv INT8_ATTENTION turns it back on).

Krea 2

krea2_turbo_t2i.json needs rgthree-comfy, installable from ComfyUI Manager. Its LoRA stack is empty by default.

| File | Folder | Size | Download | |---|---|---|---| | Krea2_Turbo_convrot_int8mixed.safetensors | diffusion_models | 12.9 GB | Winnougan/Krea-2-Base-Turbo-NVFP4-FP8-INT8 | | qwen3vl_4b_int8.safetensors | text_encoders | 4.8 GB | Winnougan/Krea-2-Base-Turbo-NVFP4-FP8-INT8 | | wan_2.1_vae.safetensors | vae | 0.25 GB | Comfy-Org/Wan_2.1_ComfyUI_repackaged |

MiniMax H3

minimax_h3_t2v_mac.json (text-to-video) and minimax_h3_i2v_mac.json (image-to-video) share the models and the turbo LoRA. MiniMax H3 needs about 40 GB of memory.

| File | Folder | Size | Download | |---|---|---|---| | minimax_h3_fl2va_pruned_int8_convrot.safetensors | diffusion_models | 21.0 GB | Comfy-Org/MiniMax-H3 | | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | text_encoders | 15.7 GB | Comfy-Org/MiniMax-H3 | | minimax_h3_video_vae_int8_convrot.safetensors | vae | 2.8 GB | Comfy-Org/MiniMax-H3 | | minimax_h3_audio_vae_fp32.safetensors | vae | 0.6 GB | Comfy-Org/MiniMax-H3 | | minimax_h3_turbo_v4_step600_pruned_comfyui.safetensors (turbo LoRA) | loras | 0.6 GB | drbaph/MiniMax-H3-Turbo-Lora-ComfyUI |

LTX 2.5

ltx2_5_t2v_mac.json (text-to-video) and ltx2_5_i2v_mac.json (image-to-video) share the models. Their optional LoRA slot is set to None: pick a LoRA and its strength on the main node to use one. The Lightricks repo is gated: accept its license on Hugging Face before downloading.

| File | Folder | Size | Download | |---|---|---|---| | ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors | diffusion_models | 21.5 GB | Lightricks/LTX-2.5 | | gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors | text_encoders | 15.4 GB | Lightricks/LTX-2.5 | | gemma4_e2b_it_int8_convrot.safetensors (prompt enhancer; the workflows won't queue without it, even with it off) | text_encoders | 5.2 GB | Comfy-Org/gemma-4 | | ltx-2.5-video-vae-bf16.safetensors | vae | 1.5 GB | Lightricks/LTX-2.5 | | ltx-2.5-audio-vae-bf16.safetensors | vae | 0.4 GB | Lightricks/LTX-2.5 | | ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors | latent_upscale_models | 1.0 GB | Lightricks/LTX-2.5 |

License

MIT; see LICENSE.