Welcome to Spirit's Lora Training! This is my first major effort to make a usable character Lora, it works relatively well, though there are improvements I would like to make in the future once I'm more experienced.
< ----- ----- ----- >
Discord - The Broken Chatbox - https://discord.gg/h3vB7S4FEw Donations! =^-^= - https://ko-fi.com/michaelpstanich
< ----- ----- ----- >
Model Name : Vert_NepSer_SD15_V1(15_e5) Experimental Name : Vert_NepSer_SD15_15_0.safetensors (My gens may have the experimental name in the data, but it's the same model, just a different name) Focus : Re-create the character Vert (Green Heart) from the Neptunia Franchise, specifically her original outfit style. SFW / NSFW : Does SFW by default, model could influence NSFW outcomes. (Training has very limited NSFW data) Training Contents : Trained on primarily official artwork and some custom in-game screenshots of the character with a primary focus on her main dress and goddess form. Training set includes some other outfits and variations to help train the character but don't seem to be replicatable in the final model.
< ----- ----- ----- >
Quick Write-up (My personal thoughts on the model and potential) : So I'm releasing this as V1.0 since I'm ok with the results, though there are some improvements that could be potentially made for a Version 2 and a potential SDXL version once I work out some quirks with SDXL training. I really struggled with getting Vert's dress to come through and had to learn a LOT to get it working without destroying the art style of models it's used with. This release version chooses what I thought to be the best middle ground between reproducing the character and avoiding overwriting a model's style. It's still not perfect but let's get to the positives and negatives.
First, the positives, I'm very happy with how this model produces both Vert's normal form and Goddess form in the same model with the potential to mix and match. I'm also happy with my keyword system that allows a good amount of freedom, HOWEVER, this keyword system means the model is harder to use than most loras you'll run into. The trade off here is the flexibility, you can use this lora for quite a lot and it's surprisingly diverse if you start experimenting. That said, let's get into the negatives. I'm not really happy with how important helper keywords are for most models and scenes, this makes things harder than they need to be, outfits are also not as consistant as I would like. I already see some potential improvements I could make to the dataset and training to fix a couple issues, but for now I'm going to move onto other model projects and potentially come back once I know a lot more about all this training nonsense.
< ----- ----- ----- >
Version 1 Usage and Information
TLDR : For Vert's nomral dress, put "VertNormal" and "VertDress" along with helper tags "blonde hair" and "very long hair" in positive prompt. Put "VertForm" and "VertArmorWhite" in negative prompt to prevent bleedthrough For Vert's Goddess form (Green Heart), put "VertForm" and "VertArmorWhite" along with helper tags "green hair" and "ponytail" in positive prompt. Put "VertNormal" and "VertDress" in negative prompt to prevent bleedthrough
Model Information (Triggers, Tags, Prompting) Triggers : No trigger was set, instead tags were used to train certain tokens Keywords Use-able : Vert (Non-transformed) : VertNormal (Some of Vert's main dress appears with this keyword as well given the heavy weighting in the training set) Vert's Main Dress : VertDress Vert (Transformed Goddess Form) : VertForm (Most of Vert's armor still shows up given the heavy weighting in the training set) Vert's Goddess Armor : VertArmorWhite Model Suggestions : I recommend using a model that uses booru style tagging as this is the tagging system used to train the dataset, doing so allows more flexibility in the output. Models without booru style taggin can be used but usage may not be nearly as flexible. I also recommend using anime or 2.5D models with anime data as they will match the training data better. Weights : 1.0 - 0.8 to strictly reproduce the character 0.6 - 0.9 when changing outfit outside the 2 main trained outfits (such as putting Vert in a swimsuit for example) 0.2 - 0.4 when using the lora to influence outfits/hair styles but limit reproducing the character
< ----- ----- ----- >
Prompting Assistance (how to get different results) : This lora was designed to allow for a good range of creativity, as such it uses various keywords/tokens to accomplish a fluid model that can be used to mix and match as well as blend. To help here are some easy guidelines for creating Vert however you see fit.
Vert in her Normal Dress : (Positive Prompt) Required Keywords : VertNormal, VertDress Recommended Helper Keywords : blonde hair, very long hair Additional Helper Keywords : blue eyes, dress, gloves, boots, large breasts (Negative Prompt) Recommended Negative Keywords : VertForm, VertArmorWhite
Verts normal dress is very RNG depending on the model, pose, and setting and may require some helper tags in some models. I only generated examples with models that worked with minimal influence using VertNormal, VertDress, blond hair, very long hair. These tags consistantly made Vert recognizable, although the outfit didn't always come through perfect it did contain all elements. If you'd like to remove elements such as the boots, gloves, ect you can try adding them to the negatives though some models don't respond as expected. Replacing elements such as gloves just requires entering in a new glove type in positive prompt with higher emphesis like 1.2 weight, though results vary by model. When altering the outfit you can reduce the weight of VertDress to help influence results, I usually set it around 0.4 weight when testing (VertNormal still includes some of the main dress data so dropping the lora weight and VertNormal keyword weight may help as well, but can make the rest of the outfit less consistant).
Vert Transformed (Green Heart Goddess) : (Positive Prompt) Required Keywords : VertForm, VertArmorWhite Recommended Helper Keywords : green hair, ponytail Additional Helper Keywords : purple eyes, very long hair, large breasts (Negative Prompt) Recommended Negative Keywords : VertNormal, VertDress Prevent Nip-Slips in some models : nipples, nude
Green Heart form comes through relatively well though the colors on the breastplate and pelvis plate may not always come through correctly. Models also tend to hate her gauntlets and as such her hands are massively difficult to get right. I used VertForm, VertArmorWhite, green hair, ponytail in positive prompt while adding VertNormal, VertDress into negatives to produce Green Heart consistantly. Without those negative prompts some of Vert's normal attire bleeds through more often, and without the green hair helper keyword models tend to still use Vert's normal hair, which I assume happens because it's more represented in the dataset. If you'd like to mix and match you can remove the negatives and adjust your prompt according to what you want to include, though results may not always be consistant
Vert but not in a trained outfit : (Positive Prompt) Required Keywords : VertNormal OR VertForm Recommended Helper Keywords : blonde hair, very long hair OR green hair, ponytail Additional Helper Keywords : very long hair, large breasts (Negative Prompt) Recommended Negative Keywords (Exclude the tag you added in positive prompt) : (VertNormal OR VertForm), VertDress, VertArmorWhite
First lower the lora's weight a bit (I usually did 0.7-0.8) then add the outfit tags VertDress, VertArmorWhite to negatives. This helps prevent the outfits from bleeding too much into whatever you're trying to create. Using the suggested positive prompts you should be able to put Vert in whatever outfit you see fit, including other lora outfits! Note however, VertNormal and VertForm still contain data pertaining to their corresponding outfits so if you experience too much bleedthrough reduce the lora weight and experiment with targetted negative prompts like collar, bow, armor, gloves, ect.
< ----- ----- ----- >
What this Lora does well : Reproduces Vert in both normal form and Goddess Form relatively well when prompted correctly Vert's hair styles come through relatively well, though they may need an assistant keyword input to show up Allows both Vert's normal dress and Goddess Form in a model for easy mix-and-match styling Easy to alter the outfit without losing character's details and traits when using helper keywords Keyword based training allows versitility in how they're implemented such as when using Regional Prompter or ControlNet Complex scenes are able to bleed through the lora, allowing a large amount of creativity Responds to basic poses relatively well (such as standing, sitting, dancing, action pose, ect)
What this Lora struggles with : Green Heart's "power symbol" don't appear by default, though you can attempt to prompt them in (trying to implement them in the data set screwed up the eyes, need to find a way to fine-tune this aspect better) Has a small impact on art style and character resolve (I struggled to get the outfits to show through without this overtraining, may be fixable in a future version with an expanded dataset and better captioning/settings) Hands and some anatomy, especially when using VertForm's gauntlets (to allow more creativity the model allows for more errors, results will heavily depend on model instead of being baked into the lora) VertNormal and VertForm Keywords for making the character have trained outfits bleed into them resulting in influence even when trying to produce other outfits, reduce lora weight to help alleviate this (I know some of this is caused by how I captioned images and may be fixed in a future version) Often requires helper keywords on various models to prevent the model overwriting the keyword weighting (I imagine this is from a limited data set and some captioning issues)
Other Quirks / Issues : Vert's size varies a bit in the data set, and this variation shows up in the Lora (Fixed by prompting your desired shape) Vert's propertions are wrong in some models for some reason (For example in RevAnimated she was often super skinny and tall, often looking stretched) Given Vert's usual pose includes crossed legs or arms results will often include these elements (Fixed by prompting poses with more detail and using negative prompts) Nude results for VertNormal may oversize her breasts (Fixed with prompting) Nude results may still include outfit elements (lowering lora weight helps with this)
< ----- ----- ----- >
AlphaTest1 Usage and Information (This information does not apply to release versions!)
Usage Information (Triggers, Tags, Prompting) Triggers : No trigger was set, instead influences individual tags allowing mix-and-match results Keywords Used : Vert's Main Dress : Vert_NepSer, Vert_Dress, vert /(neptune series/), long hair, blonde hair, hair between eyes, blue eyes, large breasts, dress, gloves, boots, Vert's Goddess Form : Vert_NepSer, Vert_Form, long hair, green hair, ponytail, purple eyes, thigh highs, elbow gloves, boots, underboob, bare shoulders, Additional Keywords (Use to influence results) : Chibi, frills, upper body, full body, Model Suggestions : I recommend using a model with booru style tagging for best results, using a non-anime style model will stil have Vert come out anime-style because the training had to be so aggressive. I personally like using, and extensively tested with SpiritMix - Soft and Cutesy (2.5D Anime). Weights : 0.6 - 1.0, 1.0 tends to work on most "3D/2.5D" models, though reducing to 0.8 may be required if backgrounds break on the model you're using. Use around 0.6 to influence outfit styles but not replicate the dress exactly, can adjust lower for varying effect.
< ----- ----- ----- >
Hey there, so like, this stuff takes a lot of work and I have many other projects, if you'd like to support my work then why not consider a donation? You can send me donations over on Ko-Fi! https://ko-fi.com/michaelpstanich
If you have any suggestions, feedback, or ideas you want to throw my way, head over to our community Discord, The Broken Chatbox! https://discord.gg/h3vB7S4FEw
- Downloads last month
- 0