Instructions to use microsoft/Mage-Flow with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use microsoft/Mage-Flow with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("microsoft/Mage-Flow", torch_dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
- DiffusionBee
issues at 2048 resolutions
From my initial testing it seems the model works perfectly at 2.5MP, anything lower or higher it starts to throw a fit. This is most likely the case with the mageFlow(non distill) one, the turbo is just horrible, so I've been using the non distill with turbo lora at low strength.

All of these were made at 2.5mp, the MageFlow + turbo lora at 0.6 strength. Sometimes you get pretty good results if you bring down the lora strength to 0.5 and use something like cfg 1.5 - 2 as well.
i was using the base model not the turbo because i hate quality degregation. that being said this is a comfy merge and recently they have been lacking in quality and a contain issues and bugs in their nodes. im going to repackage it myself and see if thats the problem
i was using the base model not the turbo because i hate quality degregation. that being said this is a comfy merge and recently they have been lacking in quality and a contain issues and bugs in their nodes. im going to repackage it myself and see if thats the problem
Not sure how you assumed I was using a turbo model when I mentioned non distill but let me clarify it once again.
Non Distill = Base = Runs with CFG and higher step count.
Distill = Turbo = Runs at CFG 1 with fewer steps.
Those images I've shared here weren't from the comfyUI implementation/ repackaged models, even before we got official support for the model, they'd shared a commit on discord and we were able to use the models from the original Microsoft repo as is. However I've compared the official models with the comfy repackaged ones and I saw no difference whatsoever. Perhaps you failed to notice what I mentioned on my comment about how the model isn't doing well above 2.5mp. Didn't it occur to you that the model wasn't trained to generate at 2k, but sure repackage it if it helps your ego.
gang you said you tested the turbo model with bad results, i didnt just pull that out my ass. and i do this for a living, theres nothing your gonna tell me that i dont already know lol
i was using comfy repackaged models so it makes sense theres some issues. also, the model DOES natively gen at 2048. do better research and lose YOUR ego gang. its literally in the readme....
"Native resolution. One checkpoint generates from 512 to 2048 on any aspect ratio, including extreme 4:1 (e.g. 512×2048, 2048×512)." 🤦♂️





