# Stepping into Flux.1 Dev this week - Is my problem the quant I am using? — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/c1d707511a54b78a

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-08-29 16:51 ingested 2026-08-30 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1w1rkl9/stepping_into_flux1_dev_this_week_is_my_problem/)

## retained text



```
I finally bit the bullet and got a reasonable GPU (3060 12GB) rather than waiting four hours for Flux.1 Dev, Steps 20 on CPU 😣.

I'm using Q4\_K\_S - (tried Q5 but I honestly couldn't see the difference and it takes longer).

People seem..plastic, unnatural. Still objects are too processed. The ages of people are sometimes way off. If I prompt for someone 60 years old, I get someone that looks mid 20-30s.

I'm working with Flux.1 Dev for the moment because I'm trying to get close to what I was able to get out of [https://draw.freeforai.com/](https://draw.freeforai.com/) \- which claims to be using the same model. Granted they are probably using the unquantized model and who knows how many steps or other parameters, they don't share meta data for me to compare. I don't know if they are using loras.

  
flux1-dev-Q4\_K\_S.gguf

t5-v1\_1-xxl-encoder-Q4\_K\_M.gguf

Steps: 20

Guidance Conditioning: 3.0

Scheduler: Simple

Sampler: Euler

  


Side note:

  
I'm setup with ComfyUI now after trying and failing with sd-cli. sd-cli kept generating 1024x1024 images that looked like upscaled 256x256. ChatGPT and I were troubleshooting trying to tweak any possible variable but in the end I tried the exact same render (models, seed, steps, guidance) and the render was 1000% better with ComfyUI. 


```



document fetched 2026-08-30 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1w1rkl9/stepping_into_flux1_dev_this_week_is_my_problem
