Sign In

FLUX.1 Dev: Sampler + Scheduler Comparison

FLUX.1 Dev: Sampler + Scheduler Comparison

Testing the outputs of FLUX.1 Dev using the same prompt, fixed seed, steps and denoise values across all combinations of samplers and schedulers. That's 27 samplers and 7 schedulers for 189 images total.

***WORKFLOW IN ATTACHMENTS***

Here's a screenshot of my global settings.

Prompt was "cinematic 18k film still, a young woman wearing glowing butterfly wings. a futuristic blue neon sign centered over here face spells Paris. She is standing in a foggy, dimly lit french palace wearing a vintage gold sequin dress. Glowing light bloom."

Here's how I connected each sampler to all the schedulers. I kept steps to 25 and denoise to 1.0.

Then I just duplicated this into large blocks with some preview grids for quick comparisons.

Here's an example of a positive outcome (green with blue check):

And a negative outcome (red):

And here's an example of an image that would probably come together with higher steps or other fixes, but I considered it a failure anyway because it's pretty similar to other pairs that were clearly better quality (red).


That said, here's two pairs that had some artifacts but had a very different result than the rest of what I was getting. dpm_fast with sgm_uniform and simple (light green with ~check):

So there's some subjectivity to the images I considered rejects. I'd encourage you to see what results you get with your own prompts and settings!

137

Comments