Many users run into friction adapting upstream weights to ComfyUI's expected folder layout and runtime variants. This repackaging removes that friction by collecting Mage-Flow diffusion weights, a matching text encoder and a VAE into a ready-to-drop ComfyUI structure, saving integration time and avoiding common file-mismatch errors.
What Sets It Apart
- Consolidated ComfyUI layout: supplies diffusion_models/, text_encoders/, and vae/ files so users don't need to move or rename multiple artifacts — so what? you can copy the folders into a ComfyUI install and start generating immediately.
- Multiple runtime variants included: bf16, int8_convrot, turbo and edit variants are present — so what? gives flexibility for quality vs. speed tradeoffs and for experimenting with smaller-int representations without re-exporting.
- Bundled text encoder and VAE: includes qwen3vl_4b_bf16.safetensors and mage_flow_vae_bf16.safetensors — so what? ensures encoder/decoder compatibility which reduces artifacts caused by mismatched components.
- Based on microsoft/Mage-Flow: repackages an upstream Microsoft model for local ComfyUI use — so what? keeps parity with the original architecture while focusing on user-facing file layout.
Who It's For and Tradeoffs
Great fit if you run ComfyUI and want a drop-in Mage-Flow setup with both performance (turbo/int8) and quality (bf16/edit) options. Look elsewhere if you need upstream model source-of-truth for research (use microsoft/Mage-Flow directly), bespoke fine-tuning workflows, or if you require documented provenance beyond a repackaged distribution. This repo emphasizes ease-of-use and file organization over providing new training or evaluation code.
