This is a LORA aims to simulate the art style of the artist known as <a target="_blank" rel="ugc" href="https://x.com/LenyavoksArts">LenyavoksArts</a>/<a target="_blank" rel="ugc" href="https://x.com/PumpknCreampi">PumpknCreampi</a> on Twitter *This artist also has a <a target="_blank" rel="ugc" href="https://www.furaffinity.net/user/pumpkincreampie/">FA account</a>. Trigger word: l3ny4v0kSty13 Recommended tag: furry, anthro Recommended LORA Strength: around 0.75 Prompting Style: e621 or danbooru tags, requires moderate amount of tags in the prompt to guide the LORA toward a decent output. Sampling method: Euler, Euler A CFG strength: 4~7 Clip skip: 2 (CivitAI on site generation will fail if you don't use 2, local generation seems to be indifferent to the Clip skip value) Post-generation Enhancement: ADetailer and HiRes.fix can help to improve the result but not always. Expected decent output success rate: NSFW: work best in 832x1216 resolution, after you found the prompt that works and in conjunction with HiRes.fix, you could get up to 60% success rate of generating a decent image. Catastrophic failure are fairly common and struggle with multiple characters. SFW: 70% success rate. However, this category of output usually doesn't require HiRes.fix to achieve that success rate, though it does still improve overall image quality. Catastrophic failures are uncommon, and subpar outputs are usually passable unless closely scrutinized. Some of the showcase images are enhanced/fixed with HiRes.fix, ADetailer, img2img or inpaint. LORA descriptions Trained on <a target="_blank" rel="ugc" href="https://civarchive.com/models/3671/yiffymix?modelVersionId=2468549">YiffyMix V63 checkpoint</a>, the images showcase here are all generated using <a target="_blank" rel="ugc" href="https://civarchive.com/models/3671/yiffymix?modelVersionId=2642671">YiffyMix V64 checkpoint</a>. The LORA yields decent results for solo character bust and half-length portraits, but the quality tends to drop as you prompt for a more complex or NSFW image. The artist’s whimsical style and character highly exaggerated expressions have a unique magic that this LORA just can't fully capture. While this LORA could delivers good enough imitations on good generation, it will never match the artist's true ingenuity, so manage your expectations (Check out the pinned post below that showcases bad image generation.) NOTE: getting good generation is a gamble, NSFW outputs tend to be lower quality. This LORA also has a fairly high failure rate when the prompt involves full-body shots, non-simple background, multiple characters or NSFW, while ADetailer and HiRes.fix can mitigate some of these issues, they unfortunately don't work every time. When generating NSFW image, the LORA tends to produce garish colors and almost always has that signature AI look. LORA known weakness: 1. Artifacting around fine details: Frequently produces artifacts around fingers, eyes, and intricate details (e.g., fishnets, teeth, fangs, shoelaces and etc). 2. Object generation bias: has a tendency to force characters to hold objects, even when unprompted. 3. Anatomical inconsistencies: Occasional anatomy failures, especially on NSFW. 4. Monochrome/grayscale bias: biased toward generating monochrome and grayscale images by default. 5. Moderate prompt dependency: Requires moderately detailed captioning/tagging in the prompt to guide the LORA toward a decent output. 6. Unprompted watermarks: Has a tendency to generate watermarks. 7. Oversaturated colors: Responds too strongly to color tags, often resulting in oversaturated tones. 8. Muted color palette: Colors can occasionally appear washed out, dull, or muddy. 9. Poor multi characters interaction: Struggles significantly with rendering correct interactions between multiple characters. 10. Weak out of knowledge handling: Fails when generating elements outside its training data: eg, attention splitting involve complex backgrounds often results in artifacting. 11. Inaccurate color fidelity: Occasionally fails to respect the character's original color palette. 12. Proportion distortion in full-body shots: Tends to shift toward a "chibi" or "short stack" art style in full-body generations. 13. Flawed hand-object interaction: Struggles to depict characters holding objects realistically or plausibly. 14. Background texture bias: Tends to generate patterned or highly textured backgrounds by default. 15. Inconsistent style fidelity: Style replication varies wildly from one generation to the next. 16. High sensitivity to monochrome tags: The LORA overreacts to tags like "monochrome" and "grayscale." 17. NSFW output tend to be lower quality: outputs almost always has that signature AI look and require HiRes.fix to look decent. 18. Inconsistent tags: eg: "5 eyes", "radio antenna", "helmet antenna" tags are inconsistent in producing the desired feature. 19. Style is not apply to background: sometime the LORA will generate realistic background. 20. Running with high strength and in conjunction with too many other LORAs will result in bad output: the only work around is lower the LORA strength, compatibility is not guarantee.

Lenyavok Style - [NoobXL] - trained on YiffyMixV63

LORA file name lenyavokStyle_yiffymix_v63Noobxl_ADAMW_V28Improvements 1. Outputs look less AI-generated, especially NSFW outputs. 2. Slightly improved prompt adherence. 3. Slightly better at generated under-represented features in the dataset. 4. Slightly more resilient from generating botched output 5. Slightly stronger style fidelity 6. Color palette choices are more similar to the art style it tries to simulate. 7. Slightly better or robust at handling out of knowledge tag. 8. Slightly more stable at higher LORA strength. 9. Less oversaturated color at higher CFG strength. Regressions 1. Character expressions are more reserved. 2. Pose are more insipid. 3. More prone to generating monochrome and greyscale image. 4. Duller color palette. 5. Less creative. Notes 1. Sometimes, <a target="_blank" rel="ugc" href="https://civitai.red/models/2706756/lenyavok-style-noobxl-trained-on-yiffymixv63?modelVersionId=3040123">V1.0</a> still outperforms this LORA for SFW bust portraits. 2. After some testing, I think I will bump the version to 1.2 considering this LORA has a non-trivial degree of improvement for NSFW generation.

Files with SHA256: 869ca97b697d2e03a1d3429964c77b0abe9766edbaffb442c9a654d8fa77c6fd