Nunchaku Fluxmania Legacy (by Spooknik)
Select Pruned model fp16 (6.3 Gb)
KREAMANIA
I don't consider Kreamania to be a truly new model, it's just a slight improvement on Legacy, injecting a tiny bit of Krea (which, in my opinion, isn't a benchmark for realism) and adding some of my recent LoRA models. It deserved the name Legacy+ more than Kreamania.
I find this version provides more detail and improved lighting. The composition remains almost identical to that of Legacy.
Note : Don't use "freckles" in the prompt
Oddity : Unexpected appearance of text
https://civarchive.com/models/1808412/wan-refine-fluxmania
https://civarchive.com/articles/17346/wan-refining-fluxmania-for-more-realism
Quantized versions of the Fluxmania Legacy model:
https://huggingface.co/belisarius/FLUX.1-dev-Fluxmania-Legacy-gguf/tree/main
Showcase images are in PNG and include metadata, prompt & workflow.
All licensing conditions for the Flux Dev model apply. You can find the details here.
Model orientation : Photography - Realism - Portraits - Artistic nude
Fluxmania III :
File 22.07 GB : unet fp16 without VAE & clip
File 11.16 GB : unet fp8 without VAE & clip
File 15.91 GB : checkpoint fp8 (not bf16) with VAE & Clip L full fp32
Settings : dpmpp_2m - sgm_uniform / cfg 3.5 / steps 25 - 30.
I recommend using this version of Clip L with the Fluxmania model
Fluxmania I & II : Unet fp8 (no VAE, no clip) / Settings : dpmpp_2m - sgm_uniform / cfg 3.5 / steps 25.
Genuine : Unet fp16 (no VAE, no clip) / Settings : Euler Simple / cfg 3.5 / steps 30.
Description
Model from a mix based on :
Genuine V1.0 fp16 (55%)
Fluxrealistic v2 fp16 (31.5%)
Fluxamnia fp8 (unpublished personal model) (13.5%)
Two Lora have been added to this mix
FAQ
Comments (27)
How'd you make it work with dpmpp_2m?
Huh. I guess it works on Simulacrum as well now, go figure.
These sampler / scheduler associations work perfectly on comfyUI :
euler / simple
uni_pc_bh2 / simple
ddim / ddim_unform
lms/sgm_uniform
dpmpp_2m / sgm_uniform
heunpp2/sgm_uniform
ipndm / linear_quadratic
uni_pc_bh2 / linear_quadratic
uni_pc / linear_quadratic
@Adel_AI
dpmpp_2m - Can't find in Forge
Damn, 3000 for download?
I'd better use another models or wait a week.
Not a week, only 3 days.
It seems huge, but the real value is only $2, which seems low to me compared to the time spent developing this model. Of course it's a merge, but it took me more than 2 months to finalize it
Rumors are that SD3.5 is far more trainable than FluxD. If so, any plans for a 3.5 in the works? Asking cause your checkpoints looks pretty sweet.
Hi, For the moment I am not considering working on SD3.5, it is certainly more malleable than Flux, but it is still very far from equaling Flux in quality. I hope the next version will be more convincing
@Adel_AI hey. Thanks for your reply. What makes Flux superior in your estimation? From what I have heard 3.5 large has far more headroom that can be developed and finetuned than FluxD. Even if it is 20% better in its current state, 3.5 could win out if there is room to improve greater than those 20% + flux headroom?
Compare SDXL base with the latest finetunes and merges, from what I have heard, it is highly unlikely we will get similar relative improvements to FluxD, and when it comes to realism, XL models can in certain cases vastly outperform FluxD. Anyhow, I do not process the competence to make such judgements.
@StreamofStars it's not a question of skill but of feeling, I like working on Flux, I go over its strengths and faults. SD 3.5 has almost the same defects as SDXL (anatomical aberrations), whatever we do, they will hamper the final result. SD 3.5 doesn't have as big a community as flux and it doesn't seem to be progressing significantly.
@StreamofStars What are you talking about? Flux with loras gives pure realism, SD 3.5 dont.
@zthrx Can you provide me with any examples?
@StreamofStars your speech is a little disconcerting, you actually give me the impression that you are commissioned by Stability.Ai to encourage creators to develop models with SD 3.5.
Taking a look at what has already been developed like LoRA with SD 3.5, the observation is overwhelming, it is dissuasive, it inspires nothing and it gives no motivation to go into this area.
But if you are convinced that SD 3.5 has potential, start developing models yourself and stop paying attention to rumors or encouraging others to do so. The one who has proven something, I think it's you, because the general discourse from the community is that SD 3.5 is far behind Flux, which is my opinion too
@Adel_AI OK, let me explain where I a, coming from:
My interest is not directed towards preserving stability AI or any other company developing proprietary or nerfed models, but towards the development of -and community engagement with models that are as open and receptive to community work as possible.
The best possible outcome, as I see it, would be if the the community, or communities, managed to find the technology and will to train our own open and 'uncensored' ( I am for banning training data based on kids being sexually abused and snuff) models, instead of (as I rightly or wrongly understand the current situation) pretty much is being allowed to do free RnD for the the companies and their proprietary Products.
If someone deems 3.5 is worse at outputting quality compared to FluxD it is not my place to try to pressure them to abandon their development of FluxD, to each their own.
Advocating for what I understand is more trainable, and less compute intense models is just me favoring the development on the path towards a higher focus on community engagement, trainability and hopefully open sourced solution of high quality.
For sure 3.5 is worse in quality, but it was my understand that this is the case as it is more more raw and undeveloped to begin with, while FluxD came out in an already pretty tuned and distilled version.
It would be fantastic if Black forest labs open sourced Flux 1.1 pro.
I agree with @StreamofStars on SDXL. To me it seems that competently-made finetunes produce more appealing images in terms of creativity, textures, lighting, composition, and overall appeal. Flux produces images that are very coherent with less mistakes, yes, but a lot of the models (bar a few), produce really plastic-y/artificial/CGI-like details and textures. And unfortunately, the ones that produce actually good textures and details, (Pixelwave and de-distilled versions like Verus Vision come to mind) don't work well with LoRAs and controlnets which make them a no-go for me.
All of these shortfalls makes SDXL still really fun and appealing to work with, for me personally. The fact that you can run it on much more modest hardware without quantization and quality loss, and with faster generation times, is an added bonus.
From what I have seen with SD 3.5, it retains the ability from SDXL/SD1.5 to produce nice textures and interesting looking images, albeit being worse than Flux in the coherency department. Something that can be improved with finetuning, I think. I feel that in the process of creating an extremely coherent model, Flux has lost the ability to create nice textures and overall appealing/interesting images. SDXL and even SD 1.5 can produce some amazingly creative images with fantastic textures and details, which I can't say the same for Flux models (bar a few that I mentioned earlier).
These are just my personal observations. I'm open to other opinions and comments. :)
@mmdd2543 @mmdd2543 @mmdd2543 Chroma might be an example of what we have been waiting for.
lodestones/Chroma · Hugging Face
It is a retrained and uncensored version of Schnell.
It is still cooking. currently at V.014
@StreamofStars Interesting! I'll keep an eye on it.
j'ai converti le modele fp8 en gguf 4ks pour le faire fonctionner sur ma machine . superbe boulot adel
Magnifique. En espérant que d'autres aussi sauront le faire, car j'ai pas envie de décliner toutes les versions possibles. Partage tes rendus pour voir un peu la qualité en GGUF
Can you give us fp16 of Version 2, fp8 seem a bit low quality than fp16
during my tests I did not notice a significant gap. I will provide the model in fp16 as soon as possible
@Adel_AI Thank you so much
@Adel_AI, By the way, this model can't not gen NSFW image right? I check your description and thought I can but so far image similar to flux dev which censored NSFW
@Wardensc I'm going to surprise you, I didn't even try to check it during my tests. I often just do artistic nudes. You may be right. I simply mentioned the artistic nude in my description and not the NSFW
Sorry I can't get the fp16 version because I threw away one of the personal intermediate models. The next versions will be edited in fp16
@Adel_AI Even in artistic nude, the nude part is very ugly and deformed
@Adel_AI Thank you
@Wardensc weird, I had not observed any anatomical deformations during my tests, but it is true that I have not done many. Flux is not a flexible model, unstable, it is capable of going from correct to ugly for unknown reason. I have encountered this situation quite often