ARTSY LITE
FLUX SCHNELL
Model Description
A very close relative of my flagship model, ARTSY FLUX.
Made with the intention of use on very limited hardware with respectable results. (The Q3 Is surprisingly good)
Artsy Lite is a careful adaptation of the dataset used to merge and finetune Artsy Flux, in order to offer similar aesthetics and performance to it's Flux.1-Dev counterpart.
However, due to the fundamental differences between Dev and Schnell, this model is not a direct copy of Artsy Flux, and may have different and unpredictable behavior by comparison. It is but an attempt to adapt the same styles, concepts, and details. As a model it is very prompt adherent, leans more towards illustration styles than Artsy Flux, and requires less descriptive prompting to produce quality results.
This Model is a sophisticated, carefully crafted merge of various self-trained LoRAs designed into introduce compositional and detail concepts into Flux.1 Schnell in order to invoke the best aesthetics and qualities that Flux has to offer. It excels at realism with diverse environments and stylistic compositions, but is also capable of 2D imagery, adding a comparable amount of detail and artistic flair. Images produced have more cohesive color grading and lighting than most base Flux.1 Schnell outputs.
I want to emphasize that this model was created for the main purpose of creating images with my personally preferred aesthetics.
BEST SETTINGS:
Sampler: Euler
Scheduler: Simple
Steps: 4
Guidance: 3.5
VERSIONS AVAILABLE:
FP8 (.safetensors)
Q8_0 (.gguf) (still uploading)
Q4_K_S (.gguf) (still uploading)
Q3_K_S (.gguf) (still uploading)
VERSIONS THAT WILL NEVER BE AVAILABLE (Do not ask):
NF4
AIO (clip_l, t5xxl and vae included)
SOME GGUF QUANTS MAY NOT BE FULLY FUNCTIONING IN FORGE UI DUE TO COMPATIBILITY ISSUES CURRENTLY BEING ADDRESSED BY ILLYASVIEL.
These models use the same clip_l, t5xxl, and ae.safetensors as Base Flux.1-Schnell.
(Images attached to each version have been generated with their respective model sizes/quants for accuracy and honesty. Though you'll be surprised at the level of detail achievable with this Q4_K_S.)
ENJOY!
Description
FAQ
Comments (24)
Wow! Great model!
Which model has better quality - fp8 or Q8? These are both pruned?
They are about the same (barely noticeable difference), but different people's configurations will get varying performance. Q8 runs better for me, but fp8 runs better for a friend of mine. It may be worth trying both to see which works best for you.
im curious whats preferred by the creator, the FP8 of Artsy Lite or the Fp8 or Artsy Flux
@TatersBarn Well, as it says in the description, they are definitely not the same model, and each one will favor different styles. So I guess my answer is that it depends on what I'm trying to make. Artsy Lite wins with 2D art and abstract composition. Arty Flux is better with finer detail with render-style and realism outputs.
@jurdn yeah i hadn't realized that lite was schnell until after i made my comment lol.
@jurdn in continued exploration of that fact, i want to report that artsy lite at 4 steps works incredibly well with loras and this is the first schnell based model i've seen work so flawlessly with dev loras
@TatersBarn That is likely due to the model actually being a merge between dev and schnell in order to bring some important artsy-relevant weights to this newer model.
Keep a look-out in the next few days for a Dev based model with some of the advantages that schnell offers, as well as some fresh styles and aesthetics. I'm thinking of calling it "Artsy Dream" right now, but that name may change.
@jurdn exciting! i look forward to it. I've got rather limited hardware resources, and both artsy lite and artsy flux both seem to run just a bit better than the other, earlier models i have of flux.d and flux.s. I have no clue as to why, but i just realized i've yet to upload a sample generation. Keep up the good work and i look forward to the next dream
This is my new favourite! Even at 4 steps it's so coherent and follows the prompt really well. Thank you!
Any plans for an nf4 version?
The Q3 is very small, and should work for anyone who would normally need an NF4, but I will not be releasing an NF4 version, sorry.
I see, I was hoping so because in my tests, nf4 safetensors runs significantly faster than gguf when both fit inside VRAM. Results from my 3060 12GB:
Q3 GGUF = 5.4 it/s,
NF4 safetensors = 2.9 it/s
Maybe you can reconsider, but I understand if this is not part of the plan. Thanks again!
I really, realy love this model and I have tried a lot- the crazy thing is I am talking about the Q4!. I trained 2 LoRA image datasets (available on CivitAI) on it from images and an embedding I made with SD1 to FLUX. Your model helped me bring to life an advanced concept I had. Thank you very much!
very impressed with the Q4 quality, for fun tested on a laptop 3060 with only 6GB of VRAM and it was still quick enough. very cool
seriously i am taken back, i think you won the Quantization lottery, this is the absolute best F.1S model in Q4! and your other Dev model isnt bad in Q4 also!, GREAT WORK
my only complant is that unlike you absolute fantastic dev model, this is a bit more photography biased, which is aboslutely fine you can see the pictures i posted are very realistic but art is a bit harder
@joehorse Thanks for the high praise! I find that Artsy Lite does at least have more illustrative bias than Artsy v1 through v3 (probably not v4).
Terms like "Illustration", "Comic book art", "2D Artwork", and "Painterly" may help you achieve less photography-leaning results.
Should guidance really be 3.5? For me 1 gives the best results (I'm using ComfyUI, if it matters).
There are TWO guidance entrypoints for Flux. In sampler you should set 1, and in CLIP encoder 3.5 or anything else
Hello! I noticed that your model Artsy Lite is licensed under Apache 2.0. However, I have a question: were any resources or code from Flux 1 Dev used in the creation of this model? If so, do any restrictions from Flux 1 Dev apply to the commercial use of your model? I’d like to clarify whether I can use this model for commercial projects. Thank you!
This project was created by merging LoRAs into Schnell.
@jurdn Thank you for your response! To clarify: were any resources, data, or components from Flux 1 Dev used in creating the LoRAs that were merged into Schnell? Do any restrictions from Flux 1 Dev apply to the commercial use of your model? I want to ensure that using this model for commercial purposes is fully permitted. Thank you!
@Eugene2077 Not at all. Feel free to use it!
@jurdn Thank You :)