Note: This model was originally posted here, and is created by Aixonlab.
Note: While the results are exceptional, potentially better than base Flux.1 Dev, beware that combining with existing Flux LoRAs does not produce fantastic results! - Ally
Lumiere Alpha
We are excited to announce Lumiere Alpha, A model focusing on improving realism without compromising prompt coherency or changing the composition completely from the original Flux.1-Dev model.
Text-to-Image
For the best results, we strongly recommend using a guidance scale of 3.0 and setting steps between 28 and 35, with Euler Beta.
Attribution notice
The FLUX.1 [dev] Model is licensed by Black Forest Labs. Inc. under the FLUX.1 [dev] Non-Commercial License. Copyright Black Forest Labs. Inc.
Our model is released under the FLUX.1 [dev] Non-Commercial License.
Description
FAQ
Comments (17)
Hey cool model thanks!
I appreciate the side by side too from dev vs this. People really should do that more often. 10/10!
I think this checkpoint don't work very well with loras, some cases is good, sometimes really bad. I don't know, maybe I'm missing something.
Correct, it does not work well with existing Flux LoRAs
I too encountered the same problem with LoRAs! ... I had generation of noise and loss of sharpness; sometimes can be also a bit "artistic" (have a look to my images posted here, in a couple of them it is quite eveident), but in general it is an issue. Tried to increase steps but not too many results even if a bit better ... :(
Excellent resource! ... quite a little different from FLUX ... people are more "real" and less glamorous and this is better! ... experiments with LoRAs give mixed results ... sometimes they give the desired effect, sometimes they generate a unavoidable noise in the image. It deserves more experimentations! (if someone else had results on this aspect, please share your opinion)
Definitely faster... I'll use it for sure! :D ... thanks to the Axionlab and theally who distributed it here.
when you get that artifact noise (or flux-grid), try bringing the shift value down, it can cook an image sometimes too. ymmv (sigh)
Still seeing cleft chins...sigh...
seems like a creative model and I like the style but often even with the recommended euler+beta 30 steps, it's still very soft and grainy. What do people recommend for tighter more detailed images?
I have the same troubles. I'm using the model in Forge with the recommended settings, but all the outputs have a very grainy appearance. Does anyone have any idea what could be causing the problem?
Ive have good luck with:
- Sampler - DEIS
- Scheduler - Simple
I upscale it and results are sharp
Euler Beta is a bit better then Simple, I also like this with the ClownShark node and the bogacki-shampine_7s Sampler w/ a beta57 scheduler, it's very slow but w/ the 8 step turbo lora it's not that bad and very very great results (with any model really but this one more so since it's going for fine details to begin with.)
I'm using dpm++2m with karras. Yes, I notice a slight bit more noise, nothing too bad which I believe is helping to give it a more realistic look.
I really like the result this model gives until as you say you introduce Lora's into the workflow...it doesn't play nice with them at all..needs to be addressed if possible to make this the best model it can be..huge upside if it can be rectified..
how does it work with art, vector, and other techniques? Thank you in advance
Your model isn't fp16, but fp8 based on size.
I wanted to dispel the notion that your model can't be mixed. I did the following:
50/50 merge with Lumiere Alpha from Aixonlab and ultrarealFineTuneV4 (Flux 1D)
plus the following (2) loras and strengths:
<lora:aidmaFLUXPro1.1-FLUX-v0.3:0.8>aidmafluxpro1.1
<lora:flux_vividizer:0.8>
You can visualize some of my results in the images below. I have referenced those with this new hybrid model.
How did you merge the models?
Excuse me can you please release the fp16 version and gguf?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.