tutututututu/Mage-Flow-Edit-Base-AllInOne

Mage-Flow-Edit-Base All-in-One · ComfyUI

A byte-preserving All-in-One Safetensors compatibility repack of Microsoft Mage-Flow-Edit-Base for ComfyUI-MageFlow.

This repository does not contain a newly trained or fine-tuned model. Transformer, Qwen3-VL text-encoder, and VAE tensors are merged into one checkpoint without changing their names, shapes, dtypes, or raw bytes.

Files

TUTU_Mage-Flow-Edit-Base-AllInOne-BF16.safetensors

• Size: 17,452,410,112 bytes (17.452 GB, 16.254 GiB)
• Format: tutu.mage-flow.all-in-one.v1
• Precision: original BF16/compatible upstream tensor dtypes
• SHA256: 30bd77f2cec6e3e6bec078dedc63920c5611730e508ef9a33fe1aaa1612c701b
• Tensor count: 1,949

Every tensor was checked against the official Diffusers source files for key, shape, dtype, byte length, and SHA256 of the raw tensor data.

Local regression result

One strict RTX 5090 / Windows / SDPA regression used the same source image, prompt, seed (230803), CFG (5.0), and 30 steps:

Package	Cold load	Edit inference	Peak allocated VRAM
Official Diffusers directory	40.10 s	22.11 s	18.251 GiB
All-in-One checkpoint	28.87 s	21.44 s	18.252 GiB

The two saved RGB outputs were pixel-exact: maximum difference 0, mean difference 0, and MSE 0.

This is a compatibility regression, not a multi-run performance benchmark. The All-in-One layout simplifies deployment and was faster to cold-load in this run; it does not reduce the model's inference-time VRAM requirement.

Installation

1. Install ComfyUI-MageFlow in ComfyUI/custom_nodes.

2. Put the Safetensors file in:

   ComfyUI/models/checkpoints/
   

3. Restart ComfyUI.

4. Add Mage Flow Edit (All-in-One) or Mage Flow Generate (All-in-One).

5. Select this checkpoint and start with sdpa, 30 steps, and CFG 5.0.

Enable cache_model when processing several queued jobs. Run Mage Flow Unload before switching to another large model.

Why a dedicated node is required

The file layout is simpler, but the architecture is still Mage Flow:

• Mage Flow DiT;
• Qwen3-VL text encoder and processor;
• Mage VAE;
• FlowMatch Euler scheduler;
• model-specific packed generation/editing pipeline.

ComfyUI's generic checkpoint loader does not currently construct this complete pipeline. The custom node hides that architecture-specific loading behind one user-facing checkpoint selector.

中文说明

这是微软 Mage-Flow-Edit-Base 的 ComfyUI 单文件兼容封装，不是重新训练的新模型。 Transformer、Qwen3-VL 文本编码器和 VAE 被合并进同一个 Safetensors，张量名称、 形状、数据类型和原始字节均未改变。

安装 ComfyUI-MageFlow 后，把模型放入 ComfyUI/models/checkpoints/，在 Mage Flow Edit (All-in-One) 节点中选择它即可。连续排队时开启 cache_model；准备切换其他大模型时运行 Mage Flow Unload 释放显存。

日本語

Microsoft Mage-Flow-Edit-Base の重みを変更せず、ComfyUI 用の単一 Safetensors にまとめた互換パッケージです。再学習・量子化・ファインチューニング は行っていません。ComfyUI-MageFlow を導入し、チェックポイントを ComfyUI/models/checkpoints/ に配置してください。

License, provenance, and responsible use

• Base model and upstream code: Microsoft Mage / Mage-Flow-Edit-Base.
• Upstream license: MIT.
• This compatibility repack: TUTU, 2026.
• No claim of original model authorship is made.

Microsoft's upstream safety screening, watermarking behavior, model card, and responsible-use guidance remain applicable. Review the official model card before use.