🏖️ Pixel Beach Beauty 👙

by Freyja Pixel 💖

version 1.1 showcase prompt reproduction from @shumaii's

Eve

version 1.0 "masterpiece, best quality, score_8, score_9, anime screenshot, 1boy, solo, muscular male, big boss, metal gear solid 3: snake eater"

🌊 General Info 🌴

These LoRAs are my first attempts at training Anima LoRAs.

Pixel Beach Beauty adds glamourous styling, cinematic angles, dynamic posing, and beautiful physique to Anima generations.

Specifically for female subjects, it emphasizes their curves and beautiful beach bodies. With all kinds of subjects and prompting, it can stylize and enhance your images, both SFW and NSFW.

Without specific tags and prompting, it will tend to put your subjects in skimpy swimwear, posing, and on a beach. It has training on details, swimsuits, posture, composition, and anatomy due to the curated training images and captioning methods. It was trained using 32 vertical images, predominantly Gelbooru tag captions with some natural English descriptions, and 1750 steps using Ostris's AI Toolkit. It was tested by community members and by myself using my Anima Twin Turbo workflow locally with ComfyUI.

📸 Purpose 🌅

version 1.0 complex prompt test

The purpose of this LoRA is to make a "beach beauty" style with a focus on anatomy, skin detail, sensual camera angles, curvy bodies, toned muscles, and sexy poses.

The ideal use case is for making spicy beach photos of your favorite waifus or husbandos, with or without other supporting LoRAs. It works best with characters (especially blue eyed blonde curvy women) which Anima is familiar and trained, like "samus aran", "whilemina, brown dust 2", "shiranui mai, the king of fighters, dead or alive", "helena douglas, dead or alive", "jane doe, zenless zone zero" or "chun-li, street fighter". It works with males, if you specify like "1boy, solo, muscular male, alucard, hellsing", "... male, geralt of rivia, the witcher series", or "... male, big boss, metal gear solid 3: snake eater".

It works well with most of the community style LoRAs I tested.

This page and the README Markdown contain all the tech details and advanced techniques!

🏄🏻‍♀️ LoRA Versions 🏄🏻‍♂️

There are two versions!

1.1 Version - 👙 Swimsuit Fashion 📸

version 1.1 with Strength 1.0 from both Pixel Beach Body and @Techannel's Sigrid

This image was generated with fixed seed on a 4 batch. Positive Prompt was "masterpiece, best quality, score_9, anime screenshot, 1girl, solo, adult woman, sigrid, sigridzzz, zenless zone zero, blush, panting, open mouth, detailed skin, sweat, large breasts, wide hips, thick thighs, barefoot, skindentation, tiny blue thong bikni with white bows, looking at viewer, facing viewer, dynamic pose, running, running toward viewer, holding ice cream cone, left leg up, legs apart, foreshortening, dutch angle, extreme low angle, depth of field, ocean spray, dark red crabs, crab swarm, swarm of monsters, water droplets, beach".

This is the recommended version, and it trained using the corrected caption and image dataset. It generally produces more reliable results, and generates a more coherent style and effect. It provides the same effects with more prompt adherence. It also produces less errors, morphs, and major anatomical failures in generations.

1.0 Version - 🌶️ Spicy Glitchcore 🥵

Lucy Kushinada by v. 1.0 with Strength 1.0. Positive prompt was "lucy, cyberpunk", and it was generated on a fixed seed on 1 image per batch.

Without the LoRA using the same generation parameters.

This is the beta version LoRA, which included some errors in the training dataset. This created some unique.... features or glitches, depending upon your desired level of erotica. This version highly favors the "hips raised" or "ass up" pose as well as any other heavily erotic composition and posing. It also tends to distort and hallucinate the anatomy of the subjects.

👀 Quick Start 💦

version 1.0 with @Nyraen07's Wilhelmina

Recommended LoRA strength: 0.9–1.0

Trigger: None

Reference model: Anima Base

Fixing your seeds and batching generations with 2-4 per run may help your results!

Recommended starting workflow:
I recommend my Anima Twin Turbo workflow with Anima Base for the checkpoint using a single K-Sampler, approximately 40 steps and CFG 4.1.

The LoRA has also been tested with dual-phase workflows, other Anima checkpoints and generation models, Anima Turbo configurations, community character LoRAs, and other style LoRAs.

Resolutions need to fall within Anima's guidelines, and I tested 768x1280, 896x1152, 1024x1024, 1024x1536, 1536x1024, and 1536x1536. I found the most stability with 1024X1536.

🌞 Basic Prompting 🏊🏻

version 1.1 with @nochekaiser881's

Jane Doe

version 1.0 Geralt of Rivia

Pixel Beach Beauty can work with extremely short tag prompts, long detailed tag prompts, hybrid tag and prose prompts, or full length natural-language descriptions.

Quality

I reccomend "masterpiece, best quality".

Score

I reccomend "score_7", "score_8", or "score_9" as a tag after your quality tags, before your style tags.

Style

"Anime screenshot" is the only "style tag" trained.

Minimal Prompt

anime screenshot, 1girl, solo, bikini, beach, sunset

Typical Prompt

masterpiece, best quality, score_7, anime screenshot, 1girl, solo, adult woman, wet skin, detailed skin, large breasts, wide hips, thick thighs, white bikini, standing, hands on own hips, looking at viewer, from side, beach, ocean waves, sunset, pink sky

Male Prompt

version 1.0 Alucard, Hellsing

masterpiece, best quality, score_7, anime screenshot, 1boy, solo, muscular male

⛱️ Recommended Tags 🦀

version 1.1 Style test with @NXMZ's

Hiromitsu Style

Character/body

adult woman, blue eyes, blonde hair, long hair, flowing hair, wet skin, detailed skin, sweat, large breasts, wide hips, thick thighs, toned female

Poses

standing, kneeling, leaning forward, front bend, hands on ground, legs apart, seiza, tiptoes, dynamic pose

Camera

front view, from side, from behind, from below, low angle shot, extreme low angle shot, cowboy shot, full body, foreshortening

Environment and Background

beach, ocean waves, sunset, pink sky, purple sky, water droplets, depth of field, warm lighting

⚠️ Minor Issues 😔

version 1.1 with @Zergkool's

Nagatoro Hayase

🛑 The LoRA has known issues with hands and feet! 🛑

If you encounter issues with morphing or bizarre anatomy, I highly recommend adding more detailed posture, composition, and gaze tags, such as "standing", "arms up", "legs apart", "looking left", "front view", "from behind", "from below", "from side", "full body", "cowboy shot", and/or "dutch angle".

🖼️ Details and Enhancements 🎨

Lucy Kushinada horizontal by version 1.1

Pixel Beach Body enhances Anima's default lighting and style to be more clear, bright, character focused, and present like an OVA beach episode. By default it enhances expression, blush, shine, sweat, muscle tone, curvaceousness, cinematic angles, clear lighting, depth of field, colorful skylines, and beach scenes. It can add detailed swimwear in your desired character's default styling, without a need for specific prompting. It enhances eroticism, sensuality, and general fanservice. It has other advanced features, detailed below!

Lucy Kushinada vertical by version 1.1

Although the current training image dataset contains no nude or explicit images, testing frequently showed increased sexualization and, in some cases, better nude anatomy and explicit poses when the LoRA was enabled. This is most evident in the 1.0 Spicy Glitchcore version. This appears to be an emergent interaction between the LoRA's learned features, the training dataset, and Anima's existing knowledge. Results vary substantially by prompt, seed, model, and other LoRAs.

This was present even with positive prompts that were random like only "hyperrealistic horror pompompurrin", which resulted in a bizarre or broken compositions without the LoRA and a blonde woman against a sunset beach scene or dark background with the LoRA.

Vertical "hyperrealistic horror pompompurrin"

with No LoRA

version 1.0

version 1.1

Horizontal "hyperrealistic horror pompompurrin"

with No LoRA

with version 1.0

with Version 1.1

🔥 No Trigger 🔫

version 1.1 with @Kayken's

Sue Storm

It doesn't have a specific trigger, it's captioned with Gelbooru tags and description words/phrases. The captions are about 90 percent gelbooru tags and 10 percent short natural English phrases separated by commas.

⛔ MAJOR ISSUES 🚫

Version 1.0 break example

Nebris Nudity and Pose test

Version 1.1 Break Example

Yor Briar Advanced Prompt Pose Test

🟡 The LoRA may have problems when used with realism LoRAs or styles! 🟡

Here are the major issues I found during my testing. I hope to address these in future versions.

The structure is:

Specific issues | What you may see | Suggested fix

Hands and feet | malformed digits or awkward contact | specify hand placement, improved negative prompt; change seed

Object interaction | warping near held objects | specify pose and object tags in prompt

Pose bias | hips-up/bent-forward poses appearing unexpectedly | explicitly state posture and composition

Muscle bias | added abdominal/limb definition | lower LoRA weight, describe softer body, improve negative prompt

Demographic bleed | blonde/blue-eyed/light-skinned traits leaking into other characters | strengthen character descriptors, add character LoRA, or adjust LoRA strengths

Female-only training | may morph bodies to feminine traits | prompt for more masculine attributes

Bikini bleed | clothing transformed toward swimsuit styling | explicitly prompt garment and reduce LoRA weight

It's Always Sunny | noticeably bright lighting or sunset effects in sky | explicitly prompt for 'night sky, shadows, heavy shadows, midnight' or similar

SFW Training | difficulty and errors in generating NSFW material | prompt for specific actions and body parts, add supporting uncensored LoRA

🏋🏻‍♀️ Training and Testing 🧪

version 1.0, male test prompt

Pixel Beach Beauty was trained locally using Ostris's AI Toolkit on an NVIDIA RTX 4070 with 12 GB VRAM.

Training Configuration

Base model: circlestone-labs/Anima-Base-v1.0-Diffusers

Dataset: 32 images

Training resolution: 768

LoRA rank: 32

LoRA alpha: 32

Batch size: 1

Gradient accumulation: 2

Training steps: 1750

Learning rate: 2e-5

Optimizer: AdamW8bit

Noise scheduler: FlowMatch

Timestep type: weighted

Training bias: style

Precision: BF16

Caption dropout: 0.10

Text encoder training: disabled

EMA: disabled

👯 Beta Testers 👯‍♀️

by M3M0G3N5 using v.1.0

Beta-testers were recruited using word-of-mouth, and they volunteered their time and teraflops.

Check out their websites and profiles (hyperlinked in their handle)!

They helped identify the error in training, verify the effective correction in the new model, and distinguish traits of images generated with LoRA. A gallery of their artwork is paired with the model showcase galleries and exhibited here! Testers reviews and social media links are listed below.

MEM0G3N5

(version 1.0)

"I like the styling it's placing upon them."

by M3M0G3N5 using v.1.0

(version 1.1)

"I wasn't sure what to expect, but even if you aren't generating beach related images, I love the style and tone this LORA put on my images. Definitely would including in your workflows if you are looking to mix up the way your base anima images look."

by M3M0G3N5 using v.1.1

by M3M0G3N5 using v.1.1

Seductress Waifu (version 1.1)

"I like it... It makes the body look so beautiful, and it made the bra and panties look cuter."

by Seductress Waifu using v. 1.1

by Seductress Waifu using v. 1.1

G Bloodforge (version 1.0)

@G_Bloodforge

"Definitely influences the style on Jumpshot to make it more anime looking. Adds some muscle toning I wasn't expecting, but that might be the seed. Running a comparison batch now.

Alright. It definitely adds a STYLE to it. Works well with most of the models I've tried it on. I was going full blast with lora strength to 1, but for the intended look? It's solid.

I think it makes Animosity look a lot better than just the base 2.0 model."

Anonymous Tester 01

(Version 1.0)

"Your Lora really likes to put them in top down bottom up pose. Cool thing I'm noticing with your Lora, if you mix it with a character tag or character LoRA, but don't include the character costume tags, your Lora will automatically take the style of their clothes and turn it into a bikini."

(Version 1.1)

"This version is nice. Kept your style but it's a lot more flexible about poses, and still works well with character Loras. If I had a set of 20 images, 10 with, and 10 without the Lora, I would be able to tell which was which."

👩🏻‍💻 Tech Details ⚙️

Motoko Kusanagi with version 1.0 for a complex prompt and composition test

Preferred Model

Anima Base is the primary and most extensively tested model.

Checkpoint Compatibility

The LoRA has also produced usable results with several other Anima-family models, including aesthetic variants, custom distillations, and merges. These received less testing than Anima Base, and their use may have differing results or require other settings.

🤸🏻‍♀️ Advanced LoRA Functions 🎞️

version 1.1 with @Hikarias's

Tifa Lockhart

It may help refine details, composition, and characters especially when using ambiguous or short prompts with subjects who are female, well known, blue eyes, blonde hair, curvy, or toned.

It may help with centering the subject in the middle of the composition and clearly defining her face, body, and features. It may also enhance composition angles by increasing focus on faces or sensual posture.

version 1.1 with @Huevoasesino2's

Academy D.VA

It naturally creates swimwear, especially bikinis, and this attire often mimics the coloring, style, and costume of a known character. This "costume" effect is present in both versions of the LoRA, and it can be enhanced by supporting the generation with a character LoRA and a few descriptive tags, such as "pink visor sunglasses" for "jubilee, marvel rivals" or "white one-piece swimsuit, purple sarong, strappy heels" for "remielle dan, seashade pas seul, zenless zone zero".

version 1.1 with @nochekaiser881's

Yor Briar

It is trained on left/right directionality using coherent images, body postures, and captions. Taking from cinematography's concepts of camera right or camera left, captions were written with a visual midline design. So as you look at the image, the right hand side of the image is "right". This training includes "looking at", "facing", and various body parts distinguished between left and right. So, in other words, "right hand" means the hand closest to the right side of the image. This function may not be present or compatible in all generations. Using prompting like "right arm up" may help improve your general composition and prompt adherence without specifically raising the rightmost arm in the composition. This effect seems to be more present when batching generations on a fixed seed between tested with 2 - 4 images per generation run.

version 1.1 with @umina's

Remielle Dan

🥇 Advanced LoRA and Anima Prompting 🤓

version 1.1 with @Konan's

Tastsumaki

It works best with characters, poses, and settings that are closer to its training. It works better with highly detailed and structured prompts following Anima's recommended structure. It works better when you "chunk" prompts sections together, so keep things that are similar close together in your prompt. For example, "long hair, blonde hair, flowing hair, red streaks in hair, hairband, blush, diamond earrings" is going to generally provide better results than "long hair, hairband, earrings, diamond earrings, flowing hair, red streaks in hair, jewelry, blush, blonde hair".

Fill-in Positive Prompt Structure

anime screenshot, 1girl, solo, adult woman, [eyes], [expression], [hair], [skin/body], [nails], [attire], [accessories], [gaze], [facing direction], [main pose/action], [arms/hands/legs], [camera angle], [shot/framing], beach, ocean waves, [depth of field], sunset, [sky color], water droplets, [lighting/weather]

Highest Captioned Positive Prompt Structure

anime screenshot, 1girl, solo, adult woman, blue eyes, blush, blonde hair, long hair, flowing hair, wet skin, detailed skin, sweat, large breasts, wide hips, thick thighs, blue fingernail polish, blue toenail polish, white bikini, hoop earrings, looking at viewer, facing viewer, [pose], [limb details], [camera angle], full body, beach, ocean waves, depth of field, sunset, pink sky, water droplets, warm lighting

Default Negative Prompt

`worst quality, low quality, score_1, score_2, score_3, artist name, bad anatomy, bad hands, extra fingers, missing fingers, fused fingers, bad feet, extra toes, fused toes, extra digits, signature, watermark`

NSFW Negative Prompt

`worst quality, low quality, score_1, score_2, score_3, artist name, bad anatomy, bad hands, extra fingers, missing fingers, fused fingers, bad feet, extra toes, fused toes, extra digits, signature, watermark, censor, pixelated, blurry'

🖥️Training Data 📈

Training image 03

Schedule and Versions

Initial training was highly successful, and the beta version of the LoRA was tested extensively. Post training and testing, AI and human analysis showed dissimilarity in the beta version of 2-3 of the training captions as compared to the rest of the data set. One of the training images required additional repair to fix an error in the background. These fixes were implemented before re-training for version 1.1. Datasets and previous LoRAs were retained with back-ups in case rollback was required.

Testing of 1.1 revealed that both version of the LoRA produced similar results in enhancing Anima generations, with 1.1 being preferred for prompt adherence and generation quality. As such, both versions were scheduled for release to the public.

Training Images

Images are 32 non-nude SFW vertial conly 768x1280 images.

The original inspiration images are from my early days in AI, used in my Sue Pixel Beach Z-Body Bikini Beach Image LoRA, and they were all originally generated using an older Wai-Illustrious 8 in Automatic1111. Before training this Anima LoRA, the entire image set was re-generated using the same prompts in ComfyUI with Wai-Illustrious 8 and ADDetailer nodes to repair some of the initial issues. Current training image limitations include female-only subjects, daytime lighting only, ambiguous hands and feet, and background morphs and artifacts. All of these issues required AI and manual edits to repair before training the current versions of the LoRA.

These limitations creates many of the known issues or difficulties within the LoRA (especially version 1.0), as many of these images have less detailed hands and feet. One of the images in the set was unable to be repaired, and it is known to have 4 fingers on both hands visible, with fingers spread and touching the ground. The image was kept in the dataset as the posture and detailing on the rest of image are fantastic, and the finger placement is physically plausible in the composition.

Please check output images for bad hands and feet during your use!

🟢 Advanced Captioning Methods 🟢

version 1.1 with @Konan's

Mount Lady

Captions were written using a complex and novel multi-step process involving multiple AI systems, human edits, and verification using an Anima image generation. Captions were analyzed by AI and errors were corrected once after initial training, for improved clarity of data to create the 1.1 Swimsuit Fashion Version.

Each caption was automatically generated by Qwen 3 VL 8B locally using Ostric's AIToolkit. The resultant captions were analyzed using an AI assistant, and manually edited by me. The automatically generated captions were extremely contradictory and flawed, and the novel validation method was introduced to produce superior captions.

Given what information is published about Anima's training, the validated training captions were designed to be 90 percent Gelbooru tags, checked and selected by a human and an AI using DraconicDesign's gitHub database (https://github.com/DraconicDragon/dbr-e621-lists-archive/tree/main/tag-lists/gelbooru). Roughly, the remaining 10 percent of the captions were designed to be natural and short English prose phrases separated by commas, and this captioning was implemented due to Anima's training including random tag drop-out. This English prose support was planned to add accuracy, precision, and diversity to the captions and LoRA functions without overwhelming Anima's Gelbooru-based training. Only English characters, commas, and very few hyphens ("-") were used in the captions, and no complex safety keywords, author tags, activation triggers, or punctuation (i.e. "@", "_", "[BREAK]", "/\", "()", or anything similar) was used in the captions.

In short summary, Anima has very flexible training data based heavily on Gelbooru tags, and this LoRA's captions were engineered to be as similar as possible while filling in any gaps with simple, short, and natural human language.

The draft validated captions were each independently tested as an image-generation prompt using base Anima without any LoRAs and fixed generation parameters. The generation used the Anima default positive prompt (masterpiece, best quality, score_7) + caption (anime screenshot, 1girl, solo, ....) as a prompt to check image, character, and pose fidelity before beginning the LoRA training runs. Using live human-in-the-loop edits, caption wording was retained, altered, or eliminated based on how reliably Anima reconstructed the intended subject, pose, composition, and background. Once the caption-as-prompt produced a reliable result which resembled the original training image, the caption was saved into the dataset. This process was done for all 32 images, with multiple drafts and checks, in the aim of adding structure to training of the LoRA. The hope is to improve prompt adherence and increase the coherence and effectiveness of the training and output images.

In other words, the captions were engineered to also act as prompts for a similar image from Anima image gens, to hopefully improve results with all types of use, LoRAs, or custom Anima models!

Captioning Tag Statistical Summary

Captions had 160 distinct exact tag strings, 1150+ total tag occurences, and a rough average of 35-37 tags per caption. The top 5 most used tags and strings are organized by semantic category and graphed below. If a section has less than 5 tags, then the training data only has the tags listed in the graph.

1. Style

| ---: | ------------------ | -------: | -------: |

| 1 | anime screenshot | 32/32 | 100% |

| 2 | warm lighting | 12/32 | 37.5% |

| 3 | soft lighting | 8/32 | 25.0% |

2. Character Description Tags

| ---: | ------------------------ | -------: | -------: |

| 1 | blonde hair | 32/32 | 100.0% |

| 2 | blue eyes | 32/32 | 100.0% |

| 3 | wet skin | 32/32 | 100.0% |

| 4 | adult woman | 31/32 | 96.9% |

| 5 | detailed skin | 31/32 | 96.9% |

3. Attire

| ---: | -------------------------------------- | -------: | -------: |

| 1 | hoop earrings | 29/32 | 90.6% |

| 2 | barefoot | 24/32 | 75.0% |

| 3 | white bikini | 19/32 | 59.4% |

| 4 | white bikini with blue straps | 10/32 | 31.2% |

| 5 | wearing white bikini with blue trim | 1/32 | 3.1% |

4. Posture, Pose, Gaze & Body Direction

| ---: | ---------------------------- | -------: | -------: |

| 1 | looking at viewer | 25/32 | 78.1% |

| 2 | facing viewer | 11/32 | 34.4% |

| 3 | facing left | 10/32 | 31.2% |

| 4 | hands on ground | 10/32 | 31.2% |

| 5 | kneeling | 10/32 | 31.2% |

5. Composition & Camera

| ---: | ------------------------ | -------: | -------: |

| 1 | 1girl | 32/32 | 100.0% |

| 2 | solo | 32/32 | 100.0% |

| 3 | depth of field | 17/32 | 53.1% |

| 4 | full body | 17/32 | 53.1% |

| 5 | low angle shot | 10/32 | 31.2% |

6. Background, Environment & Environmental Effects

| ---: | ---------------------- | -------: | -------: |

| 1 | water droplets | 32/32 | 100.0% |

| 2 | ocean waves | 30/32 | 93.8% |

| 3 | beach | 29/32 | 90.6% |

| 4 | sunset | 28/32 | 87.5% |

| 5 | pink sky | 13/32 | 40.6% |

🔎 Testing 🔬

version 1.1 with @Ringakiseki's

Akane Nanao

Image generation testing was done using ComfyUI and various other image generation platform (thank you beta testers). The preferred workflow is a single phase K-sampler with 40 steps 4.1 CFG using anima base. The LoRA was tested on both single and dual-phase K-Samplers (using my Twin Turbo Public workflow) with and without Anima turbo LoRA. Testing was a minimum of 500 image generations for each version of the LoRA. Testing included just Pixel Beach Beauty as well as a paired generation testing with 50+ separate community character and style LoRAs from CivitAI.com and CivitAI.red. Character and style LoRAs were purposefully chosen to diversify testing between SFW and NSFW subjects and prompts. For final phases of testing, example prompts from other LoRA creators were taken from CivitAI to test compatibility with alternate prompting styles, and this produced favorable results in most cases.

A POTENTIAL INTERACTION WITH REALISM LORAS WAS FOUND DURING TESTING. I would not recommend mixing these two styles strongly. The output images tended towards uncanny valley and strange compositions, and this was MORE present with version 1.0.

📃 LoRA Test List📎

version 1.1 with @JudgeNaegi's

Kurakashiki Reina

I am publishing a small gallery for each of the tested LoRAs on my CivitAI. This may take some time to organize and fully post. Here is a condensed list of the creators and their LoRAs, hyperlinked in a similar manner as the image credits. I'm listing Style LoRAs first, as I tested fewer of them, and they will take up less characters and space.

All LoRAs are presented in alphabetical order according to the files in my ComfyUI. In other words, most ordering is arbitrary outside of the Top 3 prompt adherence, which are also ranked arbitrarily.

Some LoRAs may have been updated since I downloaded them for testing. I used the links from the rgthree Power LoRA Loader to hyperlink to the exact LoRA model I used in the workflow.

To test compatibility with Pixel Beach Beauty with the Style LoRAs, I sorted for prompt adherence and ranked the top 3, and I did not sort for aesthetics or appearance beyond creating a coherent image. The Style LoRAs listed below are a selection of the large pool that were tested, and the listed ones I deemed to have "high prompt adherence", meaning the subject and posing were correct according to my prompt and directionality with Pixel Beach Beauty.

To test compatibility with Character LoRAs I tried making one image using each LoRA, a tag for the character, and one or two additional LoRA or franchise tags. I wrote simple and complex tag-based prompts including tags from my training captions. After I achieved a desirable and non-morphed image, I ran a 4 batch generation. I saved and published the best results. GENERALLY, THE FEWER images in the galleries, THE BETTER, as those LoRAs had the easiest time in creating my vision and desired image before the batch generation.

Style LoRAs

Top 3 prompt adherence

Portrat: Flashy by @artyclaw

Stickybuns 3D Style by @gothwomenenjoyer

Takeda Hiromitsu Style by @NXMZ

Others tested

Hi-Res / Aesthetic Boost by @circlestone_labs

Greg Rutkowski Style by @circlestone_labs

3d cg style by @itshoney

Porn Star Ass by @Ringakiseki

Thebrobine Style by @guy90

Character LoRAs

Emma Frost and Sue Storm and Black Cat and Jubilee and Jubilee Midnight and by @Kayken

Hell Knight Ingrid and Assumi and Jaina Proudmore and Sally Whitemane and Lara Croft by @Enigmata

Nebris and Whilhelmina by @Nyraen07

Evelyn and Mai Shiranui and Misery and Sorceress: Dragon Crown and by @DeskGrenade

Burnice White by @Hoseki

Kurashiki Reina and Lily Ramses and Miyajima Tsubaki by @JudgeNaegi

Jessica Rabbit and Mitsuri Kanroji and Queen Marika and Seras Victoria and ShadowHeart and Mount Lady and Tatsumaki and by @Konan

Jill Valentine and Juno and DVA Academy by @Huevoasesino2

Lucy Kushinada by @BySp1rit

Marin Kitagawa by @MetalchromeX

Meiko Shiraki by @EasonnLi

Meryl Silverburg by @Jackkarate

Nagatoro Hayase by @Zergkool

Olga Discordia by @KeMiliUs

Promeia by @Sansenskiy

Rieri Bishop by @NotAFinancialAdvisor

KOS-MOS by @Sqquirtle0007

Akane Nanao by @Ringakiseki

Eve by @shumaii

Loen Celebrity Bunny by @noire_v

2b by @w63725310859

Marciana Marine by @rift1000

Remielle Dan Swimsuit and Remielle Dan White by @umina

Sigrid by @Techannel

Tifa Lockhart by @Hikarias

Yor Briar and Yoruichi and Yoruichi TYBW and Jane Doe by @nochekaiser881

📅 Future updates and testing ⌨️

version 1.1 with @Nyraen07's

Nebris

Partial training testing

I have multiple prototype LoRAs created during training which may produce better results or have superior prompt adherence. These need thorough testing to see if they have any use.

Training Mergers

If any partial LoRAs prove useful, I may attempt merging them, adding training to them, or even doing further image generation testing.

Other Models

Due to the Gelbooru tag heavy captions, training may be possible on Z-Image, SDXL, or Krea2. If so, this would be done before before altering the dataset.

Image dataset updates

NSFW and Explicit Images

Adding in between 8 - 16 NSFW, nude, and solo explicit images in the same style to increase ability to generate NSFW material

Anatomical and Posing Images

Adding in between 8 - 16 high anatomical accuracy and neutral pose images on various colored background to balance posing, lighting, musculature, anatomical consistency, and default skin coloration. I may add in a few high hand and foot focus images to assist with these known gaps.

Diversification of Ethnicity, Gender, Eyes, and Hair

I would like to add in multiple diverse characters to the Anatomical and Posing Image set to help even out the training of the LoRA. This would be up to 24 additional images added in the hopes to improve the general functions of the LoRA.

Updated Captioning

Captioning New Images

This is the largest time sink, and it may be prohibitive in what I am able to effectively train and test.

Adjusting Current Captions

AI assisted checks and human analysis of existing captions may find additional gaps or errors for correction and retraining.

Advanced Verification

Future caption validation will continue using base Anima without the LoRA as the primary verifier. The previous Pixel Beach Beauty release may additionally be tested as a secondary diagnostic for backward compatibility and behavioral comparison.

Thanks for your time, energy, and support! 💖

version 1.0 using a rewritten prompt from @Enigmata's showcase

Lara Croft

Files with SHA256: 6c4fb688ffd0b1455b2e963392a3c1a10ea04416d710acd7a0408b72fb150b4c