Elves can work. I got a dwarf too but needed photoshop to make face rounder and body shorter.

@Anashel hope to try it soon

@Anashel I remind you just in case

@Anashel Many Thanks, still throws conversion error.

Converting to Diffusers ... Traceback (most recent call last): File "/content/convertodiff.py", line 1130, in <module> convert(args) File "/content/convertodiff.py", line 1081, in convert text_encoder, vae, unet = load_models_from_stable_diffusion_checkpoint(v2_model, args.model_to_load) File "/content/convertodiff.py", line 846, in load_models_from_stable_diffusion_checkpoint info = unet.load_state_dict(converted_unet_checkpoint) File "/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py", line 1671, in load_state_dict raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format( RuntimeError: Error(s) in loading state_dict for UNet2DConditionModel: Missing key(s) in state_dict: "up_blocks.0.upsamplers.0.conv.weight", "up_blocks.0.upsamplers.0.conv.bias", "up_blocks.1.upsamplers.0.conv.weight", "up_blocks.1.upsamplers.0.conv.bias", "up_blocks.2.upsamplers.0.conv.weight", "up_blocks.2.upsamplers.0.conv.bias". Unexpected key(s) in state_dict: "up_blocks.0.attentions.2.conv.bias", "up_blocks.0.attentions.2.conv.weight". Conversion error

I tried to use your huggingface path to the avoid conversion difussers step but throws the following message.

Resolving deltas: 100% (89/89), done. From https://huggingface.co/Anashel/rpg [new branch] main -> origin/main From https://huggingface.co/Anashel/rpg branch main -> FETCH_HEAD error: Sparse checkout leaves no entry on working directory

The huggingface repo should to have a specific format??.

Regards!

if you still havnt figured out . let me know.
its the colab version thats outdated. i have fixed mined

@usamakenway271 Hi, the problem still persists. If I use the Anashel/rpg path from Huggingface, it is not recognized. The issue is that the pickletensor version is not available now, so I cannot use the direct Civitai path to test.

@Anashel 

Thank you. Does the V4 includes Hasan blend or other NSFW models?

Works fine on my gtx 1080 (8 Gb VRAM) and RTX 2060 mobile (6 Gb VRAM). If you can run SD 1.5 you can run this model.

I have 1660s 6G, works fine

@Lancer4658 mine is 1060

@Anashel 请小心可能的陷阱。如签订商业条款或劳务合同请务必明确内容。

By the way,your work is excellent. It deserve five stars. Love comes from China

Yeahh, works fine.

@itsMedic Could you maybe help me get it working? I have it installed and have used some of the example prompts from the PDF and all I'm getting are garbage sketches...anyone know what I'm doing wrong?

Yes!

https://www.youtube.com/watch?v=m-UVVY_syP0&t=439s

thanks

I'm curious what setting(s) I could possibly modify within automatic to get more similar results to the results I fell in love with using the RPG model on leonardo. Thanks in advance for any insights!

Yes, I have read your excellent user guide. I was just curious if you knew why two seperate applications using your model give such different results, but if you haven't tried leonardo, I appreciate that you wouldn't be privy to that. Thank you for the response and thank you so much again for your work!!!!!

Funny you said it because I came here after using RPG on Leo and my results are similar/better to the ones on Leo

@Anashel when you go to the community feed, keep clicking the images until you find one made with RPG, scroll down a bit and you'll see images that were made with the model, on the right there'll be a button saying "View More" and you'll see all the pictures made with it

In case you don't get a response from the author, I happen to believe it was not. From what I understand, this model was created primarily with 3D content from a game developer's own library. They made these models to rapidly prototype new character designs and decided to share with the community. But, take that for what it is worth, as I'm a 3rd party to all that.

@zackstone 
Thank you very much for the information.

@Anashel 
I'm impressed by the results.
Only 5% was enough to stabilize the model in v2? Usually, it takes much more...
Did you continue the training of v3 from v2, and v4 from v3?

@Anashel 
I feel that it's a shame that people do not share more about the process of training and model merging, it's important information that can be beneficial to everyone. Obviously, nobody should reveal secrets that he does not want to reveal.

Many people probably make a lot of the same mistakes and keep inventing the wheel each time.

@Anashel 
Yes, 21 iterations is a lot.

@Anashel , dude thanks for all that effort, I really love this model, I am new to this but the model is on point

@Anashel Thanks so much for the forthright info, I'm so impressed by your model and plan to use it as an ingredient in a merge for base training, the composition and concepts are just too good. It's tough to find models that aren't muddled and mysterious so again, deep appreciation for your work!

I'm no expert but it might be due to the resolution size you are using. I read somewhere that if you increase the resolution too much, the models can "get confused" and fill in the extra space by cloning stuff. You could try reducing it a bit (512x768) to begin with and then upscale them. (I think the models are trained at certain resolutions and when you go too far above those resolutions, you start to get double heads etc)

Yup, I'm in love with this model too. Thanks for creating it!

seems it using the combination of the two that is at fault, well that is what I think.

https://civitai.com/models/90365

@Valerius_SPQR thank you very much

You are probably using a bad resolution. Try to stick to 512*768 for example for the first generations.

@Anashel I just use automatic1111. I use xformers and medvram. But apparently people say you lose quality with one of them but im not too sure. As for plugins I use controlnet but they weren't used on my test generations. As for the settings im just copying what was seen in the guide so close to those settings yeah

@Azuki900 can you explain what you mean by mushy, low definition?

@lukeovermind Basically yeah. Just to clarify its not a resolution issue as I always render on high resolutions. But every time I generate the faces just look all blotchy and patchy and my generations dont look anything close to the examples

@Azuki900 "I always render on high resolutions" - that is your problem, use max res of 768x768 or more likely 512x768 etc. If you like what you get use tile diffusion to upscale it or lock the seed lower cfg and use highres fix. Never just generate in high resolution. Also for faces you often need to add extra prompts like "ultra detail face" etc

@bitzupa Well the thing is I've always used Hires fix and I dont go over 768 x 768. But even then im not getting good results

@Azuki900 ohhh i see, i thought you made mistake many new users do and rendered in fullhd or 4k from start ;) Then try using easynegative2 and similar negative loras, they do work wonders. If you do already then i simply don't know what can be the reason. Maybe send me PM with prompts and setting and seed so i can check it out of curiosity.

@Anashel Sweet

@Anashel This is just fantastic!

@Anashel , fantastic news! saw the red light next to the notification icon one here and hoped it was an update! there goes my sleep!

@Anashel thanks!. i used vae-ft-mse-840000-ema-pruned. and turned out pretty good

Could you give a prompt and the parameters you use, so we can check?

Nice.

@Anashel For sure, i'll be happy to do it! :)

Indian dot thing (bindi) I'd imagine. Put "indian", "india", "hindi", "bindi" and the like in the negative prompt.