ComfyUI FP8 checkpoints
This repository provides ComfyUI-ready checkpoints derived from the official Microsoft Mage-Flow-Edit and Mage-Flow-Edit-Turbo weights.
Mage-Flow is not currently supported by official ComfyUI. To load and run these checkpoints, use the custom nodes and workflows from starsFriday/ComfyUI-MageFlow.
| File | Component | Recommended settings |
|---|---|---|
mage_flow_edit_fp8_e4m3fn_scaled.safetensors |
Mage-Flow-Edit transformer, scaled FP8 weights | 30 steps, CFG 5.0, Euler, simple |
mage_flow_edit_turbo_fp8_e4m3fn_scaled.safetensors |
Mage-Flow-Edit-Turbo transformer, scaled FP8 weights | 4 steps, CFG 1.0, Euler, simple |
qwen3vl_4b_fp8_scaled.safetensors |
Qwen3-VL 4B text encoder, scaled FP8 weights | Use the Mage-Flow prompt template |
flux2-vae.safetensors |
Flux2 VAE | Recommended for the current ComfyUI workflow |
mage_vae_bf16.safetensors |
Original Mage-VAE in BF16 | Included for compatibility; see the warning below |
The FP8 transformer files use scaled FP8 weight storage with
BF16/full-precision matrix multiplication. In the ComfyUI model loader, select
weight_dtype = bf16; activation-quantized fast FP8 matrix multiplication can
make the four-step Turbo checkpoint collapse into structured noise.
[!WARNING] The official Mage VAE currently produces garbled or noise-like decoded images in this ComfyUI workflow. Use
flux2-vae.safetensorsfor normal image decoding.mage_vae_bf16.safetensorsis retained as the original BF16 reference checkpoint.
ComfyUI custom nodes
Use the workflows and loader/conditioning nodes from:
Install the node package under ComfyUI/custom_nodes and place the downloaded
files as follows:
ComfyUI/
├── custom_nodes/
│ └── ComfyUI-MageFlow/
└── models/
├── diffusion_models/
│ ├── mage_flow_edit_fp8_e4m3fn_scaled.safetensors
│ └── mage_flow_edit_turbo_fp8_e4m3fn_scaled.safetensors
├── text_encoders/
│ └── qwen3vl_4b_fp8_scaled.safetensors
└── vae/
├── flux2-vae.safetensors
└── mage_vae_bf16.safetensors
Restart ComfyUI after installing or replacing the files so its node and model caches are refreshed.