LoRa model of Murdoch from the visual novel The Smoke Room by Goodnight Fellowship
This LoRa was trained with Illustrious model, but all sample images were made with NoobAI both models can be found here in Civitai
If you are having problems I recommend copying the prompts of the samples and using them as a base
Trigger Word: "Murdoch"
Reinforcement Prompts: "green eyes,"
Description
FAQ
Comments (5)
After, looking at the v1 vs the v3 version of this model. The v3 has some definite improvements like adding more detail, better lighting, I think the backgrounds don't look out of focus blurring people tend to do which looks more visually pleasing for art. When you could see the art rendered more clearly. The v1 is still great on its own. Showing how the model has progressed over time.
I think because, both versions have distinct traits. Both are worth keeping for different takes of the same character. Since, the og has a less digital paint look . While the v3 version has more of a modern digital paint look. Kinda like old vs art vs remastered vn in some games.
With some editing. The models could be used in some before and after interactive renders which seem like it could be worth trying out. It's great to see that the original and latest model has benefits I think would please different people. For different reasons.
Thanks for the comment 🤗but actually, I don't like V1 at all, (assuming that you at least know how Murdoch looks in TSR) I don't feel like the one in V1 is him, more like another fox doing cosplay if you know what I mean, things like the colors, body parts, and proportions are terrible in V1 and the same happens with V2, and while V3 is not perfect it definitely looks more accurate and could be even more with the right combination of artists tags
@Orion_12Â I think the issue is your training using stable diffussion 1.5 instead of Stable difussion 2.0 or sdxl pony. Which would train the model to look much more like him. Are you using sprite dumps or screen shots? Because of you use then renpy .py scrip dumper for the sprites you might get better model results.
I don't think using artist tags would help much. That sounds like it would make the art even more inaccurate to the visual novel. If you dont have the sprite dumps I could do a quick dump and link you the files if that might help.
Don't worry, this one and all my VN models are made with sprites and CGs extracted from the games, thanks for the offer though, aside from that, I don't really know much about SD 2.0 or XL, I once tried using Pony Diffusion XL to train Loras and the results were just blurry images, and to use it simply as a base model I didn't like it either, I feel you don't have much control over the style of the images (without using Loras). maybe I'm just using it wrong or something but in version 6 none of the artist tags worked for me ("by meesh" for example) and using it only with captions or tags didn't give me satisfactory results, they were very varied outputs, from cubism to anime, passing by a weird combination that seemed like Zoruik + BNA_V5, so yeah, I am not interested in doing SDXL Loras right now, maybe in the future when there are more furry alternatives to Pony Diffusion
@Orion_12Â Sorry if this late but, I went ahead and did a redump of all the sprites and used smilingwolf/wd-convnect-taggeerv3 to auto tag and do prompts. Not sure, it this would help you any
https://www.mediafire.com/file/r7huas3w0vrsayj/TSR_sprites.zip/file
Did you use sdxl chkpt or lora? Because, with sdxl pony lora models. I found that doing character specific models trained with a character name does do a better job than sd1.5 with the benefits of a hd res training vs the 512p limitation sd1.5 is limited to which can make sd 1.5 art look less trained than sd2.0 and sdxl lora models.
I've done characters like link (wolf form), and some other characters with the sd1.5 than sdxl pony. An it does seem like sdxl pony does train better and create better furry art when used with the sdxl pony model. Though, on my computer using nmkd stable difussion gui. I think the sd1.5 lora models work better than how civit ai handles the rendering back end. Which could just be a server thing. Since, free users seem to use less ai options than paid users from what I've noticed.
So, it could just be how the site handles lora models or something.