Flux quality subjectively regressed in beta v3

[Bersaelor]

What are you trying to build?

I'm working to generate playing cards using Flux and SD.

I first discussed build issues in this issue.
When I am testing the Flux model in Beta 27 compared to previous versions I feel like the quality has come down..

Where are the current docs or utilities unclear?

Beta 3 seems to objectively lead to lower quality outputs then previous iterations.

Example prompt:

a playful cartoon illustration of a single [Ant], [green] background, no scenery, bold clean outlines, flat shading'

Each card has the same prompt just subject and background color changed.

Expected improvement

Any ideas what to do to get better flux outputs? This seems much worse than what I would expect given how good Flux usually is.

Additional context

No response

[stikves]

#115

One option is disabling quantization of the text encoder model (Qwen3), as the images themselves seem to be not losing any quality, but only not adhering to the prompt as well.

(I've checked the guidanceScale, which is the default 1.0, and did not see any other changes that could affect this)

Can you give this a try and let me know?

[stikves]

Thinking back, we quantize the text embedding model to able to fit flux pipeline in iPad RAM. It comes with some quality cost, but should not be as much as what is reported here.

So, closing that PR for now.

[Bersaelor]

I'll gladly give it another try in the next beta where changes are made, just give me a ping.

[Bersaelor]

I mean many of the Flux outputs are hilarious in their weirdness, but I haven't had much luck actually generating something that looks like an actual picture of drawing of the subject

[kevchengcodes]

Hello, we just merged #120 that may help with image quality. Can you try again on the tip of main?