johnsor/Mage-Flow-Edit-Turbo-GGUF

Mage-Flow-Edit-Turbo, Q6_K

A GGUF quantization of Mage-Flow-Edit-Turbo's diffusion transformer for stable-diffusion.cpp: 3.15 GB, down from 8.23 GB at bf16.

Mage-Flow-Edit-Turbo is a 4B native-resolution multimodal diffusion transformer that edits an image from an instruction — hand it a photograph and ask for a scene containing it. It runs in 4 steps, which is what makes it practical on a CPU.

Why Q6_K and not something smaller

Below 6 bits this transformer does not survive stable-diffusion.cpp's uniform quantizer. Converted from the same bf16 source and run with identical settings:

Type	Size	Result
bf16 (source)	8.23 GB	clean
Q8_0	4.08 GB	clean
Q6_K	3.15 GB	clean — this file
Q5_K	2.65 GB	a wavy lattice over flat areas, worst at 1024
Q4_K	2.33 GB	heavy corduroy ribbing over the whole frame
Q4_0	2.17 GB	no image at all, flat pink ribbing

Q6_K is the floor. An earlier revision of this card recommended Q5_K, on the strength of a busy 512 px edit where it does look clean — but on a flat background it lays down a faint regular lattice, and at 1024 px that becomes obvious. The artifact is systematic texture, not graceful degradation, and it gets worse with every bit you remove. If you need 4-bit, this is not the model for it.

mage-flow-edit-turbo-Q5_K.gguf is still in this repository for anyone who wants to see the artifact, or who would rather have the 0.5 GB. The others are reproducible in a couple of minutes with the command below.

Files it needs

This is the transformer only. It also wants:

• Text encoder — Qwen3VL-4B-Instruct-Q4_K_M.gguf (2.33 GB) and mmproj-Qwen3VL-4B-Instruct-Q8_0.gguf (0.42 GB), both from Qwen/Qwen3-VL-4B-Instruct-GGUF. The mmproj is what lets it look at the reference image; without it a reference is ignored.
• VAE — mage_flow_vae_bf16.safetensors (0.32 GB) from Comfy-Org/Mage-Flow.

About 6.2 GB all in.

Running it

sd-cli \
  --diffusion-model mage-flow-edit-turbo-Q6_K.gguf \
  --llm Qwen3VL-4B-Instruct-Q4_K_M.gguf \
  --llm_vision mmproj-Qwen3VL-4B-Instruct-Q8_0.gguf \
  --vae mage_flow_vae_bf16.safetensors \
  -r reference.png \
  -p "a photograph of this dog wearing a red knitted scarf, sitting in a snowy park" \
  --steps 4 --cfg-scale 1.0 --sampling-method euler \
  -W 512 -H 512 --diffusion-fa --offload-to-cpu -o out.png

4 steps and --cfg-scale 1 are what the Turbo checkpoint wants; the base checkpoint uses 30 steps. Multiple -r flags pass several reference images.

Example

The reference, and the output of the command above at 512x512 — 99 seconds end to end on a Ryzen 7 4800H, CPU only. These two images were made with the Q5_K file, at the size and subject where it holds up; Q6_K is a little slower and does not lay down the lattice on flat backgrounds.

Reference	Result

How this was made

sd-cli -M convert \
  --diffusion-model mage_flow_edit_turbo_bf16.safetensors \
  -o mage-flow-edit-turbo-Q6_K.gguf --type q6_K

Source: diffusion_models/mage_flow_edit_turbo_bf16.safetensors from Comfy-Org/Mage-Flow. stable-diffusion.cpp at revision c92d73c408515c94beef32161bb5960764fde7a0. Conversion takes about 75 seconds.

Licence and credit

MIT, following microsoft/Mage. The model is Microsoft's; this repository contributes nothing but a quantization. Note that the microsoft/Mage-Flow and microsoft/Mage-Flow-Edit-Turbo repositories on the Hub are not currently reachable, so Comfy-Org's repackage is the source used here.

Quantized for use in offgrid.