PJ0_KREA
REAL1SM_V3_FP8/FP16
Most likely, this is the latest version of our joint model with my friend https://civarchive.com/user/Triple_Headed_Monkey . Styles and details have been improved in this version. As a result, we achieved what we had planned at the beginning of our work - realism and art in one model.
For the style I'm using a node ComfyUI_MileHighStyler - Flux Mile High Beta | Flux Other | Civitai
For realism and art styles use combination of samplers - Euler/Simpler, for anime and illustration use - Euler_ancestral/normal
svdq-int4-project0_v40ArtReal1sm
Special thanks @theunlikely for the work done on quantification!
And thanks to JIB (J1B Creator Profile | Civitai) for his WF and clarifications on using this type of model:
For using this model you need to install the nunchaku project following the instructions there. and the nunchaku ComfyUI custom nodes to get it to work. Download and unzip the archive from Civitai to: \comfyui.git\app\models\diffusion_models\int4-project0_v40ArtRealism-svdq\ This is the Nunchaku workflow I am using: https://civarchive.com/models/617562 Thanks a lot to theunlikely for running the Quantization which takes an H100 GPU 6 hours!
Project0_REAL1SM_FP16 v2
In this version, the details have been worked out, the background, the styles are more pronounced, the landscapes and nature have been trained.
DO NOT POST NSFW CONTENT. If I see this, I will delete such content.
Project0_REAL1SM_FP16
In this version, more attention is paid to realism, and a small merge is made with a small part of the TruBass2 Tensor.Art model by Triple_Headed_Monkey.
In order to keep up to date with the latest news in the AI world, I invite you to join our server AI Revolution https://discord.gg/hUKgjWwN
DO NOT POST NSFW CONTENT. If I see this, I will delete such content.
Ver4_Art_Realism
DO NOT POST NSFW CONTENT. If I see this, I will delete such content.
Improved background detail, realism, and some styles. This version of the model contains a small part of the TruBass ver 2 model published on TA
It's not exactly a realistic model, but the realism is a little better than in version 3 and it's very good for an art model.
v3 art
This version is a transitional to the art version. Many art styles have been worked out, realism has been slightly reduced, and the model still does a good job with realism.
v2
Has more realism, emotion and quality.
You can download the quantized versions of the model from the link ND911/Project0_GGUF at main
Thanks kjs195 Creator Profile | Civitai for the quantized versions of the model Project0_v2
V1
This model is an experiment to improve the basic DEV model. I worked on the model together with my friend https://civarchive.com/user/Triple_Headed_Monkey The model is experimental and I hesitated to publish it for a long time, but since many people ask for it, I do it. The model gives decent results only in the bf16 format. When you try to use it in fp8 format, the result becomes much worse. This is the final version (the model is under development) and I will try to fix it.
Thanks to OldCelticSoul for providing the opportunity to convert to GGUF format. https://huggingface.co/ND911https://huggingface.co/ND911
Thanks kjs195 Creator Profile | Civitai for the quantized versions of the model Project0_v1. You can find the rest of the versions here ND911/Project0_GGUF at main
Description
Project0_REAL1SM_FP16
In this version, more attention is paid to realism, and a small merge is made with a small part of the TruBass2 model.
FAQ
Comments (23)
Quants for Project0_REAL1SM https://huggingface.co/ND911/Project0_REAL1SM_FP16_GGUF another fun model!
@kjs195 thank you so much
thank you
Thanks! Any chance for FP8? GGUF is really nice for VRAM savings, but twice as slow and Flux is already pretty slow on anything other than RTX 4080/4090. Thanks in advance!
@mmdd2543 I can publish the FP8 version, but it will be of lower quality. and i need some time for it
@kjs195 Thank you so much for your help
Will try 4. Using Draw Things, I haven't been able to get any of your previous Flux models working yet. The posted drawings make me very enthusiastic...
- edit: YES! nice results, added dynamics and better artistic interpretations. very prompt-friendly :-)
Your comments seem to indicate that you're very much anti-NSFW. Is that a fair assumption? Does that mean you have no intention of improving the "anatomy" of men and women in your model?
I don't want to see such images in the gallery (NSFW). If someone likes it, you can generate similar content, but please enjoy it yourself. About improving the anatomy? This is not the main task at the moment. Everything is done consistently and according to plan. If do it now, I'm afraid it will disrupt all the work done before. I hope I won't offend anyone with this.
@speach1sdef178 Not offended at all. Just good information to know for different use cases.
Она станет моей любимой моделью. Спасибо, круто!
Hello guys,
I love the model but I can't find the : Flux_Finish_StylesStyler node that you using
if you can give me the github repo please 🙏🙏
@djodjo Good afternoon! This node is part of the MileHigh node. I have uploaded an archive with a json file as Training Data, you can download it and put it in the directory "...\ComfyUI\custom_nodes\comfyui_milehighstyler\data\FLUX_Finish_Styles\"
Then you need to restart ComfyUI.
Also here is a link to the main MileHigh node https://civitai.com/models/119246 ?modelVersionId=933860
OMG! THANK YOU SO MUCH!!!
I've been looking for a checkpoint like this for so long. This is pretty much the best checkpoint I've ever seen. This is better than Flux 1.1 Pro Ultra to me. I can't believe this exists, more people need to find out about it. It's that good. These look like real high quality pictures. I've been playing with character LORAs and without character LORAs I trained and the pics are unbelievable. These pictures are so scary real that they would fool people.
I'm using this without LORAs to enhance the images. You can just use the checkpoint by itself and you'll get great images. This is pretty much all you need going forward.
I thank you for releasing to us, this is a huge secret model that needs more exposure.
@BlairBrooksBB Thank you for your kind words. I am very glad that you like the result. and I think this is not the last checkpoint. I and Triple Headed Monkey are continuing to work on our series of models!
@speach1sdef178 You're welcome and I thank you again for releasing this to us. I'm glad I stumbled upon this as I've spent the last 10 months looking for the best checkpoint that produced a good balance between amateur photography and cinematic photorealistic images, and this does just that, as I'm kind of addicted to it right now. LOL! That is so great to hear you guys are going to keep working on a new checkpoint, as that's very encouraging. I will start following you, so I can get updated on whenever it comes out.
This does the best job that I've ever seen with full body images. I would even like to see even more improvements made there, if possible.
@BlairBrooksBB we will improve this as much as possible, and maybe even more. Thank you for support! :)
@speach1sdef178 You're welcome and great to hear to hear you guys improving. I know it's tough to get the full body photos from a distance right, faces can still come out wonky at times, but this is the closest I've seen it coming out pretty good, I just have to generate a few times to get a really good one.
can u share your training setup especially learning rate?. I've trained on other checkpoints with aitoolkit but result are all bad but all the loras I've train on base flux dev are good. it feels like all the other checkpoints has too strong effect on loras. subjects resemblance always lacks.
Hi,which version do you usually recommend for illustration style designs?
@Esdeath_Sama Hi! For illustrations, I think the best solution would be version 3 (FP8) or the Real1sm (FP16) version, and for the best result of illustrations or animation styles, I recommend using a combination of euler_ancestral/normal samplers or dpmpp_2s_ancestral/ddim_uniform.
@speach1sdef178 thank you so much you're the best
do you have any sampler recommendations for realism as well? (I am usually using Euler/Simple)
@Esdeath_Sama Hi! thank you fo your words))) For realism, I use euler/simple and 30-36 steps, and I like to use 1408*1024px image resolution.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.