# IkariDev > My personal site: machine learning, anime, and software, plus the latest posts from the blog. ## Pages ### Home https://ikari.sillytilly.org/index.md

Welcome to my site!

This site was inactive and not maintained for about 2 years. It got rebuilt on 15 September 2026, read more about that in the rebuild post.


Moin, i am Ikari. I am a AI/ML enthusiast from germany (she/her, he/him also works), in my blog, i talk about all kinds of things, here some examples: - AI/ML - Drawing - Technology - Archiving - Anime - Games [SFW and NSFW] - And lots more! I love all kinds of anime, from shitty isekai to romance and stuff like Clevatess. Very cute and very gay! Also draw sometimes now, you can check that out on [twitter](https://x.com/IkariDev_tw). Been deep into modding [My Dystopian Robot Girlfriend](https://incontinentcell.itch.io/factorial-omega) with C#/Unity lately too. I love robots stamp

If you want to know what i am currently up to you can check out my [blog](/blog/). ### About https://ikari.sillytilly.org/about.md ## Hey, i am IkariDev. ### I am a developer from germany (she/her, he/him also works) doing various things, such as: - Machine learning [ Large Language Models, ~~Stable diffusion~~(I distanced myself from this, i hate seeing slop everywhere now), Upscaling, other small stuff ] - Watching anime - Writing software/scripts - Drawing - Modding games [ C#, Unity, currently modding MDRG via MelonLoader ] ### My main programming languages i use: - C# - Python ### Also know a bit of: - Rust - TypeScript/JavaScript ### Programming languages i want to learn more: - Rust ## What i like: - Anime - Edgy stuff (I mean, look at my website LMFAO) Self expression is not a crime - Open-Source - [My Dystopian Robot Girlfriend](https://incontinentcell.itch.io/factorial-omega) ~ best NSFW game, no contest I love robots stamp - The Coffin of Andy and Leyley ~ best SFW game ## What i dislike: - OpenAI - People who don't read Readme files before asking questions ### You can also check me out on these places: - [Github](https://github.com/IkariDevGIT) - [Civitai](https://civitai.com/user/ikaridev) - [Twitter](https://x.com/IkariDev_tw) ~ art stuff Nothing much to say about me, but if you really want to know more about me you can check out my [blog](/blog/). I know you are but what am I? --- ## Contact You can currently contact me on these services: - Discord [ username: ikaridev ] Shut up and leave me alone ### Projects https://ikari.sillytilly.org/projects.md #### MoanMod Status: Active Project started: 20 February 2026 Added to site: 2026-09-16T01:44:08.000Z A NSFW [MelonLoader mod](https://github.com/IkariDevGIT/MDRGMoanMod) for [My Dystopian Robot Girlfriend](https://incontinentcell.itch.io/factorial-omega) that adds moaning, breathing and facial expressions during sex scenes, tied to her pleasure/lust/sympathy stats. Everything is tweakable live and in-game via my good friend Echo's [Mod Settings Menu](https://github.com/Echo5Dev/MDRG-ModSettingsMenu). - [GitHub](https://github.com/IkariDevGIT/MDRGMoanMod) - [Official itch.io thread](https://itch.io/t/6004461/-moaning-mod-moaning-and-expressions-for-jun-by-ikaridev)
For more info see the blog post. /mdrg-moan-mod/ #### UnityExplorer for MDRG Status: Active Project started: 1 September 2026 Added to site: 2026-09-16T01:44:08.000Z A [UnityExplorer](https://github.com/yukieiji/UnityExplorer) fork that works with [My Dystopian Robot Girlfriend](https://incontinentcell.itch.io/factorial-omega). - [GitHub](https://github.com/IkariDevGIT/MDRGUnityExplorer) - [MDRGUniverseLib - GitHub](https://github.com/IkariDevGIT/MDRGUniverseLib) #### Legacy projects These projects wont be updated as of 15 September 2026. They're mostly outdated and only kept around for the history. ##### Athena-13b Status: Legacy Project started: 30 August 2023 Added to site: 2026-09-16T01:44:08.000Z A [Large language model](https://huggingface.co/IkariDev/Athena-v1) i made 30 august 2023. **History:** It was was once deemed first place back in 2023-08-31 on the [ayumi_erp_rating](https://rentry.org/ayumi_erp_rating). Athena lost it's place on 2023-09-15, it was beaten by [Undi95/MLewdBoros-LRPSGPT-2Char-13B-GGUF](https://huggingface.co/Undi95/MLewdBoros-LRPSGPT-2Char-13B-GGUF). ##### Athena-v3 and Athena-v2 Status: Legacy Project started: 23 September 2023 Added to site: 2026-09-16T01:44:08.000Z 2 merges i made.. idk when i made it, i forgot. [Athena-v3](https://huggingface.co/IkariDev/Athena-v3) and [Athena-v2](https://huggingface.co/IkariDev/Athena-v2). **History athena v3:** Idk, i forgot.. is a great model tho! **History athena v2:** Idk, i forgot.. was a good(kinda) model. Update 19.10.2023: There is also a Athena-v4 now, but no one cares about v4 because its not as good as v2 and v3 lol. ##### Echidna-13b model series Status: Legacy Project started: 21 October 2023 Added to site: 2026-09-16T01:44:08.000Z - [Echidna-13b-v0.1](https://huggingface.co/NeverSleep/Echidna-13b-v0.1) - [Echidna-13b-v0.2](https://huggingface.co/NeverSleep/Echidna-13b-v0.2) - [Echidna-13b-v0.3](https://huggingface.co/NeverSleep/Echidna-13b-v0.3)
For more info see the blog post or model cards. /legacy/echidna-13b-collab-with-undi/ ##### MinervaAI Status: Legacy Project started: Q3 2023 Added to site: 2026-09-16T01:44:08.000Z [MinervaAI](https://huggingface.co/MinervaAI) is a team i am in, we are making LLM models. Find more info on the page. **Aesir:** [Aesir](https://rentry.org/ayumi_erp_rating) is a LLM model + Dataset we are currently working on. Want more info? Check out the blog post. /legacy/minervaai-aesir/ ##### Noromaid model series Status: Legacy Project started: 13 November 2023 Added to site: 2026-09-16T01:44:08.000Z **Noromaid 0.2 [ EXPERIMENTAL ]** - [Noromaid-13b-v0.2](https://huggingface.co/NeverSleep/Noromaid-13b-v0.2) - [Noromaid-7b-v0.2](https://huggingface.co/NeverSleep/Noromaid-7b-v0.2) - Spin-Off's - [FlatOrcamaid-13b-v0.2](https://huggingface.co/NeverSleep/FlatOrcamaid-13b-v0.2) **Noromaid 0.1.1** - [Noromaid-13b-v0.1.1](https://huggingface.co/NeverSleep/Noromaid-13b-v0.1.1) - [Noromaid-7b-v0.1.1](https://huggingface.co/NeverSleep/Noromaid-7b-v0.1.1) - [Noromaid-20b-v0.1.1](https://huggingface.co/NeverSleep/Noromaid-20b-v0.1.1)
For more info see the blog post or model cards. /legacy/noromaid-model-series/ ##### Other models Status: Legacy Project started: 12 January 2023 Added to site: 2026-09-16T01:44:08.000Z - [HuggingFace - IkariDev](https://huggingface.co/IkariDev) - [HuggingFace - NeverSleep](https://huggingface.co/NeverSleep) ### Resources https://ikari.sillytilly.org/resources.md Here you can find various resources i am using/i used/i suggest using. ## ML ### LLM #### Models - [HuggingFace - NeverSleep](https://huggingface.co/NeverSleep) ~ Org by me and Undi. — added 22 Sept 2026 - [HuggingFace - Undi / Undi's Kofi](https://huggingface.co/Undi95) ~ A shit-ton of high quality merged models. — added 22 Sept 2026 #### Backends - [koboldcpp](https://github.com/LostRuins/koboldcpp) ~ My new main backend for LLM stuff, only for llama.cpp. — added 22 Sept 2026 - [Llama.cpp](https://github.com/ggerganov/llama.cpp) ~ Inference of LLaMA model in pure C/C++. It can run on cpu+gpu, good option if other model loaders require too much vram. — added 22 Sept 2026 - [Tabby](https://github.com/theroyallab/tabbyAPI) ~ Backend for exllamav2 and exllamav3. — added 22 Sept 2026 - [Sonar](https://github.com/dphnAI/sonar) ~ The official backend engine for PygmalionAI, used to be called Aphrodite. — added 22 Sept 2026 - [OobaBooga-textgen webui](https://github.com/oobabooga/text-generation-webui) ~ (old)My main backend for LLM stuff. — added 22 Sept 2026 #### Frontends - [Sillytavern](https://github.com/SillyTavern/SillyTavern) ~ A really good RP/ERP ui. Based on TavernAI, but sillytavern is much better. — added 22 Sept 2026 - [Vali's ChatterUI](https://github.com/Vali-98/ChatterUI) ~ Basically SillyTavern but hosted on your phone. — added 22 Sept 2026 #### Character cards - [Chub](https://www.chub.ai) ~ A site full of chars you can download. — added 22 Sept 2026 - [Character Tools](https://character-tools.srjuggernaut.dev) ~ A collection of tools for creating and managing characters for AI Roleplay. — added 22 Sept 2026 - [Chatbots Webring](https://chatbots.neocities.org/) ~ Webring for charcard websites. — added 22 Sept 2026 #### Various discords - [NeverSleep](https://discord.gg/mGU9CVb6vV) ~ NeverSleep server, by me and undi. — added 22 Sept 2026 - [SillyTavern](https://discord.gg/sillytavern) ~ SillyTavern server. — added 22 Sept 2026 - [TheBlock](https://discord.gg/theblokeai) ~ TheBloke server. — added 22 Sept 2026 - [KoboldAI](https://discord.gg/X3WmRM9Npn) ~ KoboldAI server. — added 22 Sept 2026 #### News - [HuggingFace Daily Papers](https://huggingface.co/papers) ~ Daily papers. — added 22 Sept 2026 - [4chan Lmg](https://boards.4chan.org/g/catalog#s=lmg) ~ Sometimes maybe good sometimes maybe sheet. — added 22 Sept 2026 - [LocalLLaMA](https://www.reddit.com/r/LocalLLaMA/) ~ Subreddit to discuss about Llama, the large language model created by Meta AI. Often really hangs behind news. — added 22 Sept 2026 #### Learn - [Neural Networks: Zero to Hero](https://www.youtube.com/playlist?list=PLAqhIrjkxbuWI23v9cThsA9GvCAUhRvKZ) ~ Learning the basics of how AI works. — added 22 Sept 2026 - [Attention is all you need](https://www.youtube.com/watch?v=bCz4OMemCcA) ~ A complete explanation of all the layers of a Transformer Model: Multi-Head Self-Attention, Positional Encoding, including all the matrix multiplications and a complete description of the training and inference process. — added 22 Sept 2026 - [annotated-transformer](https://nlp.seas.harvard.edu/annotated-transformer/) ~ Idk what to write here, but someone said i should link it here. — added 22 Sept 2026 - [Kalo's sampler post](https://www.reddit.com/r/LocalLLaMA/comments/17vonjo/your_settings_are_probably_hurting_your_model_why/) ~ Your sampler settings are probably hurting your model, why? — added 22 Sept 2026 #### Model creation/Dataset Creation - [MergeKit](https://github.com/cg123/mergekit) ~ Suite of merging scripts. With those scripts you can merge LLM models with a variety of different merging methods. — added 22 Sept 2026 - [Augmentoolkit](https://github.com/e-p-armstrong/augmentoolkit/tree/master) ~ Generate multi-turn training data, about any subject, using Open Source LLMs. — added 22 Sept 2026 #### Datasets - [ERP/RP and erotica rawdata collection](https://rentry.org/qib8f) ~ A collection of datasets. — added 22 Sept 2026 ### Stable Diffusion - [vladmandic](https://github.com/vladmandic/automatic) ~ A fork of AUTOMATIC1111 which does things differently. Try which fits best for you. — added 22 Sept 2026 - [ComfyUI](https://github.com/comfyanonymous/ComfyUI) ~ Yet another powerful SD UI, here you have way more control than in vlad, or auto1111. ComfyUI will let you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. — added 22 Sept 2026 - [kohya_ss](https://github.com/bmaltais/kohya_ss) ~ A tool for training sd Loras/Locon/whatever. — added 22 Sept 2026 - [CivitAi](https://civitai.com/) ~ A place where you can get various loras/models/whatever. — added 22 Sept 2026 ## Web development ### Neocities Neocities type of sites. - [sadgrl.online](https://sadgrl.online/) ~ Bunch of stuff neocities related. — added 22 Sept 2026 - [y2k](https://y2k.neocities.org/) ~ Like the last one but not sorted. — added 22 Sept 2026 - [gifcities](https://gifcities.org/) ~ Archive of neocities-like gifs you can use on your website. — added 22 Sept 2026 - [ezgif](https://ezgif.com) ~ A really good online tool for modifying gifs. — added 22 Sept 2026 ## Archiving - [ArchiveBox](https://github.com/ArchiveBox/ArchiveBox) ~ Self hosted archiving, saves whole pages as html/pdf/screenshots so they dont just vanish. I host it myself too. — added 24 Sept 2026 - [gallery-dl](https://codeberg.org/mikf/gallery-dl) ~ Command line downloader for image galleries and collections from a ton of sites. — added 24 Sept 2026
## Something missing?

Resources that went down or got abandoned move to the inactive resources page. They wont be maintained, that does not mean they are not worth a look.

### Inactive resources https://ikari.sillytilly.org/resources/inactive.md Resources that went down, got abandoned or are not worth using anymore, kept around so the links do not just vanish. ## ML ### LLM #### Models - [HuggingFace - TheBloke](https://huggingface.co/TheBloke) ~ here you can find various GGUF quantizations. (inactive: No new models for years.) — added 22 Sept 2026 - [HuggingFace - AzureBlack](https://huggingface.co/AzureBlack) ~ The TheBloke of exl2 quants. (inactive: No new models for years.) — added 22 Sept 2026 - [Gryphe/MythoMax-L2-13b](https://huggingface.co/Gryphe/MythoMax-L2-13b) ~ Was once the best model available for RP/ERP, some still consider it best. (inactive: Was once the best, not anymore.) — added 22 Sept 2026 #### Model ratings - [Snombler](https://snombler.neocities.org/logs) ~ Model ratings. (inactive: The site is down.) — added 22 Sept 2026 - [BestERP](https://besterp.ai/s/models) ~ Nice site for model ratings. (inactive: The site is down, not sure if it is coming back.) — added 22 Sept 2026 - [Ayumi ERP rating](https://rentry.co/ayumi_erp_rating) ~ An Erotic Roleplay model rating by weicon. (inactive: Not active anymore.) — added 22 Sept 2026 #### Tools - [OSGA](https://github.com/ElliottDyson/OSGA) ~ OSGA(Open-Source Generative Agents) is a community-driven initiative that builds upon the foundation of 'Generative Agents', for local LLM models. It is based on this Paper (https://arxiv.org/abs/2304.03442) (Video: https://www.youtube.com/watch?v=nKCJ3BMUy1s). (inactive: Not sure what happened to it.) — added 22 Sept 2026 #### Learn - [Ask max](http://ask-max.neocities.org/) ~ If you don't know how something works, ask max! (inactive: Not active anymore.) — added 22 Sept 2026 - [Thoughts guide (kinda outdated)](https://rentry.org/thoughts-guide) ~ A guide by me on how you can implement characters thoughts into sillytavern. (inactive: Not active anymore.) — added 22 Sept 2026 #### Model creation/Dataset Creation - [MergeMonster](https://github.com/Gryphe/MergeMonster) ~ An unsupervised merging algorithm for Transformers-based language models, using a list of phrases (both good and bad) and a fully automated strategy that strives to decrease (or increase) the probability of these phrases occurring in the final merge. (inactive: Not maintained in 2 years.) — added 22 Sept 2026 - [Quantization script](https://github.com/DocShotgun/LLM-notebooks/) ~ Bunch of Colab/Laggle notebooks for quanting models, applying loras and bunch of other stuff. (inactive: Not maintained in 3 years.) — added 22 Sept 2026 - [airoboros](https://github.com/jondurbin/airoboros) ~ using large language models to fine-tune large language models. (inactive: Not maintained in 2 years.) — added 22 Sept 2026 #### Datasets - [Dampf's list](https://rentry.org/datasets-llm) ~ A collection of datasets. (inactive: Not maintained.) — added 22 Sept 2026 - [PygmalionAI/PIPPA](https://huggingface.co/datasets/PygmalionAI/PIPPA) ~ Official Pygmalion dataset. (inactive: Not active anymore.) — added 22 Sept 2026 - [Xilabs/pippa-alpaca](https://huggingface.co/datasets/Xilabs/pippa-alpaca) ~ Pygmalion PIPPA dataset with alpaca format instead of metharme. (inactive: Not active anymore.) — added 22 Sept 2026 ### Stable Diffusion - [AUTOMATIC1111](https://github.com/AUTOMATIC1111/stable-diffusion-webui) ~ A really good ui. (inactive: Not maintained anymore.) — added 22 Sept 2026 ## Blog posts Posts marked (legacy) were migrated from the old site and wont be updated. That does not mean they are not worth reading. ### How i watch everything in 3d https://ikari.sillytilly.org/blog/how-i-watch-everything-in-3d.md Published: 2026-09-25T02:16:38.000Z Updated: 2026-10-16T13:27:21.000Z ( ˶ˆ ᗜ ˆ˶ ) I watch anime, movies and series in a pretty weird way. Nobody asked about it, but i wanna talk about it anyway, maybe theres something in here worth stealing for your own setup. TLDR; everything gets turned into 3d in real time and thrown onto a huge virtual screen, while i lay in bed and a subwoofer at the foot of it shakes the bed frame. ## The setup The thing everything else is built around is [iw3](https://github.com/nagadomi/nunif/blob/master/iw3/README.md), which is part of nagadomi's nunif project. You give it flat 2d video, it gives you SBS (side by side) 3d back. Nothing has to be made for 3d, thats the entire point, it just works on whatever you throw at it. Im using [desktop mode](https://github.com/nagadomi/nunif/blob/master/iw3/docs/desktop.md), which converts whatever is on your screen live instead of converting a file first, and i use the local viewer, not the wifi streaming thing it can also do. From there its two screens and the headset: - screen 2 has my browser with the anime or series playing - iw3 grabs that screen and converts it to 3d in real time - screen 1 has the local iw3 viewer with the SBS output on it - the quest 3 sees screen 1 through Virtual Desktop, an app that streams your pc to the headset over wifi, in desktop mode with the SBS option on so it reads the two halves as one 3d image So the video path is screen 2 -> iw3 -> screen 1 -> quest 3. For the audio, the quest is connected to my surround setup over bluetooth, so it goes pc -> quest -> surround. Sounds like that should be a latency nightmare but its actually fine considering iw3 also adds a bit of latency, it basically cancels itself out. That surround setup is 4 speakers, a center and a subwoofer. Pretty normal so far. The less normal part is that i have a second subwoofer, which is pushed right against the wooden frame at the foot of my bed. Wood carries low end really well, so the whole bed turns into part of the speaker. Explosions, swords hitting each other, doors getting kicked in, someone getting launched through a wall, you feel all of it through the mattress. Its incredibly immersive! For the headset itself, the stock quest 3 strap is not made for laying around for hours, so i use a [BOBOVR M3 Pro](https://www.bobovr.com/products/bobovr-m3-pro) instead. Its a halo strap, which means the weight sits on top of your head instead of getting pressed into your face. Because of that i also took the facecover off the quest completely, so nothing touches my face, no sweat and no pressure marks and stuff like that. The not so nice thing about that is that light gets in from everywhere once the facecover is gone, which is why i only watch stuff at night. Im also plugged into the outlet the whole time, so battery is just not something i have to think about. Then i lay down, turn the fan on, put the headset on, and watch the thing. Night person What comes out of all that: - the screen is as big as i want it. not "big tv" big, more like sitting in a cinema big - everything is in 3d! - the audio has actual physical impact behind it because of the bed subwoofer - im half sitting half laying, with the back of the headstrap leaning against the wall behind my bed, so my neck doesnt have to hold the headset up at all The 3d is what surprises people the most when i explain it. Youd expect a model guessing depth out of a flat image to look pretty bad, and it really doesnt. Its also not as 3d as an actual 3d movie, but thats a good thing, you have to balance it so its immersive without going over the top or being hard on your eyes. Ive adjusted the values so much by now that i can confidently say im incredibly happy with it. Scenes get actual depth instead of stuff flying at your face. Backgrounds sit back, characters sit in front of them, rooms feel like actual rooms. Once you get used to it, going back to flat video feels kinda wrong. ## The technical details iw3 runs a depth estimation model on every single frame. The model looks at the picture and estimates how far away everything in it is, mostly from perspective and from how objects overlap each other. You can pick between a bunch of depth models, ZoeDepth, Depth-Anything v1 and v2, Depth Pro and Video-Depth-Anything, and they all trade quality against speed a bit differently. For the second eye it just takes the frame you already have and shifts the pixels sideways based on the depth map. Close stuff shifts a lot and far stuff barely moves, which is basically what your two eyes see anyway. iw3 has a few different methods for that, the naive fast ones ghost around the edges of objects and the ML based ones handle edges way better and even fill in the gaps that show up behind them. Theres a bunch of settings that change how the 3d actually looks, stuff like divergence, convergence, IPD (interpupillary distance) offset and foreground scale. Im not gonna go through all of them here, the [iw3 readme](https://github.com/nagadomi/nunif/blob/master/iw3/README.md) explains them way better than i could and has example pictures for most of them. Performance was a real problem at first. Everything runs on a 3090, and it took a good amount of tuning before the iw3 output landed where it is now, around 20-24 fps. Written down that sounds rough, i know, but cinema has been running at 24 fps since the late 1920s. It started as a [compromise between the speeds theaters were already using](https://en.wikipedia.org/wiki/Frame_rate) when sound came in, but people connect it with cinema now and its still nice to look at. Its also doing a lot per frame, a depth model and then the warp into a second view, all in real time while the pc is also decoding video and streaming a whole desktop to a headset. The docs default desktop mode to 15 fps, so 20-24 is doing alright. For anime it also matters way less than you'd think, since a lot of it isnt animated on ones (a new drawing every frame) to begin with. Fast action scenes you notice it, everything else you really dont. And even then it doesnt bother me much, the rest of the setup easily makes up for a few missing frames. Also, if anyone is curious, i dont run the original nunif. I use [my own fork](https://github.com/IkariDevGIT/nunif/tree/opinionated) on the opinionated branch, which is behind upstream by a bit but has two things i wanted for myself. One is fullscreen display mode switching in the local viewer, so hitting F11 switches the monitor the viewer sits on to a bigger display mode and puts it back when you leave fullscreen. The other is a fix for how iw3.desktop hands frames over when you capture with wc_cuda. Nothing huge, but i use both every single day. That branch isnt really meant for anyone else, its just my own thing that i keep around because i need it. You can absolutely try it anyway, or just steal the idea and build it into your own setup, all welcome lol. And this is what i actually start it with: ```sh python -X faulthandler -m iw3.desktop --local-viewer-fullscreen-display-mode --monitor-index 1 --disable-draw-cursor --synthetic-view both --compile --foreground-scale 0.5 --screenshot wc_cuda --full-sbs --stream-fps 60 --depth-model VDA_Stream_B --scene-detect --divergence 1 --convergence 2 --edge-dilation 2 --method row_flow_v3 --stream-height 1080 --ema-decay 0.99 --ema-buffer 2 --ema-normalize --stream-quality 100 --local-viewer ``` ## Future stuff Ive tried a few more things on top of all this, most of them didnt make it. ### Frame generation This was the big one. The idea was to stop running depth estimation on every single frame and instead reuse or interpolate the depth from the frame before it, which frees up a lot of performance. Theres a [whole issue about it](https://github.com/nagadomi/nunif/issues/619) where me and nagadomi went back and forth on it. My early testing looked really promising, i went from around 23 fps to 42 with frameskip 1, and even frameskip 4 didnt look awful at around 70. Doing it properly is the hard part though, you need the depth map itself moved along with the motion, and most frame interpolation just hands you a middle frame instead, which isnt the same thing at all. Between that and a few reasons of my own, i failed and scrapped the idea. #### A simpler idea Writing this post got me thinking about it again. What if the skipped frames just blend the depth map from before and the one after, nothing clever, no optical flow. It would probably look bad at something like frameskip 6, but at frameskip 4 it might be completely fine. The nice thing is that only the depth is a guess. The image you actually look at is still the real frame, so the risk is wrong parallax and not a made up picture. Calm scenes should be fine. Moving edges are where it would wobble, since an outline would sit at some in between depth for a frame or two, and edges are what your eyes are pickiest about. Maybe i should pick this back up at some point though. Ive grown a lot as a dev since i last touched it. ### Live upscaling and restoration Cleaning up h264 artifacts and grain while its playing. I decided against that one pretty fast. I dont wanna sit there and watch slopified anime, and stacking that on top of iw3 doesnt look viable performance wise either. ### Image enhancing This is the one i still actually like. Filters and color correction, so stuff like pushing bloom a bit harder or giving a scene more atmospheric depth. My test implementation had performance problems and im not good at color correction, so i scrapped it too, but the idea itself still seems very doable. ## "Thats not how its meant to be watched" Some people are gonna say im overdoing it, or that this isnt how the creators wanted their work to be seen. The "as intended" version doesnt exist at home anyway. Stuff gets rescaled, recompressed and reframed, and most people are watching on a phone, or on a tv with motion smoothing still switched on, in a bright room, through whatever speakers came in the box. If we want to be strict about intent then nobody is watching anything correctly. I want to watch Saga of Tanya the Evil like im standing in the middle of it, and thats reason enough. ## Should you do this Its a pile of things duct taped together that were never designed to work as one system, and it somehow ended up being the comfiest way i've ever watched anything. Old shows that never got a 3d release, random stuff in a browser tab, doesnt matter, it all comes out the other end in 3d on a cinema sized screen. Not all of it is great though. iw3 isnt super stable, at least on my machine it likes to crash after a couple of hours and i have to restart it, which i can thankfully do from inside the headset. Everything also has to be connected and running before i can start watching anything. And the quest 3 does its own weird things sometimes, a controller not connecting, Virtual Desktop refusing to connect, stuff like that. So it comes down to whether youre willing to go the mile, set all of this up and buy the hardware for it. I got lucky and did it the other way around, i already had all of this stuff sitting here anyway, so i just optimized the whole thing around what i already had. If you already own a headset, iw3 is free and you can just try it. The bed subwoofer is optional, but i do recommend it. ## Afterword This got way longer than i planned, i sat down to write a few paragraphs about how i watch anime. Big thanks to nagadomi for iw3, none of this works without it, and for putting up with me in that frameskip issue. If you end up building something like this, or you actually try the subwoofer against the bed thing, id like to hear about it. My contact stuff is on my [about page](/about/). ### Undi95 is stepping away for a while (repost) https://ikari.sillytilly.org/blog/repost/undi-stepping-away.md Reposted from: https://huggingface.co/posts/Undi95/589980321681772 Published: 2026-09-16T01:44:08.000Z Updated: 2026-09-20T21:01:43.000Z We have been through a lot of finetunes and merges together over the years, Noromaid, Echidna, and so on. Before he left for the time being he still dropped Hanami as a last gift, a simple, phone friendly frontend with memory management instead. Go check it out. ### Remade the website. https://ikari.sillytilly.org/blog/site-rebuild-2026.md Published: 2026-09-16T01:44:08.000Z Hellu! ⸜( ´ ꒳ ` )⸝ Its been like 2 and a half years since the last post. dont kill me pls. The old site was one big jquery single page app: index.html swapping divs around for every "page", a jquery tooltip plugin thing, DHTML popup windows, a loading screen to hide the incredibly bad loading times, per-post vote buttons (You could press them as often as you want XD), and every single blog post crammed into one blog.html. Images were just gifs/pngs, no webp, and the visitor counter was some old janky widget, not goatcounter. Anyway, i remade the whole site. It runs on Astro now instead of one giant html, so actual separate pages instead of everything living inside index.html. Blog posts are real markdown files with their own pages now too, instead of all being crammed into one blog.html. All not really interesting for non-devs. The look stayed mostly the same tho, same colors, same header and same sidebar, i love the look and will never change it I love my site (I had many people calling my site eye cancer but i dont get it.). About/Projects/Resources content is basically unchanged too. I made the site better too, dropped the SD Gallery, Apps, and Experiences pages since they werent worth porting over. Old posts got migrated over too, so the history's still here if anyone wants to look through it. The roughly 2 year gap was mostly burnout, from AI work and making models and everything around it. If you want to know what i've been doing since, the newer posts cover it. Currently as of writing this, the site has a content and history gap of about 2 years. I've also been stepping back from the LLM/finetuning scene lately. Base models got good enough for RP on their own that finetuning isnt as necessary anymore, and money problems and general life stuff played a part too. ### Noromaid model series (legacy) https://ikari.sillytilly.org/blog/legacy/noromaid-model-series.md Published: 2023-12-19T20:51:00.000Z Updated: 2024-02-07T18:20:00.000Z [NeverSleep Discord](https://discord.gg/w6C8mys35E) Name comes from:
Noro = No\_robots(dataset),
Maid = Minerva AI Dataset MinervaAI Dataset referring to: Aesir --- ### Noromaid 0.1.1 - [Noromaid-13b-v0.1.1](https://huggingface.co/NeverSleep/Noromaid-13b-v0.1.1) - [Noromaid-7b-v0.1.1](https://huggingface.co/NeverSleep/Noromaid-7b-v0.1.1) - [Noromaid-20b-v0.1.1](https://huggingface.co/NeverSleep/Noromaid-20b-v0.1.1) Yet another merge with undi, came out 16.11.2023, main part of it a dataset by [MinervaAI](https://huggingface.co/MinervaAI) named aesir, other part is a [modified version of No_Robots](https://huggingface.co/datasets/Doctor-Shotgun/no-robots-sharegpt). ### Noromaid 0.2 [ EXPERIMENTAL ] - [Noromaid-13b-v0.2](https://huggingface.co/NeverSleep/Noromaid-13b-v0.2) - [Noromaid-7b-v0.2](https://huggingface.co/NeverSleep/Noromaid-7b-v0.2) As seen in the title, this model is pretty experimental! This time we used [MergeMonster](https://github.com/Gryphe/MergeMonster), with many bad_phrases. We also used another [MinervaAI](https://huggingface.co/MinervaAI) dataset. The 13b was published on 16.12.2023, the 7b was released on 21.12.2023, further B's will be released when we get more feedback. ### Noromaid 0.3 - [Noromaid-13b-v0.3](https://huggingface.co/NeverSleep/Noromaid-13b-v0.3) Noromaid 0.3 has: new datasets, normal alpaca Instruct(with instruction, input, and response header) instead of our modified alpaca. It is a complete retrain based on the same model as 0.1.1 and 0.1. ### Noromaid 0.4 - [Noromaid-7b-v0.4](https://huggingface.co/NeverSleepHistorical/Noromaid-7B-0.4) - [Noromaid-7b-v0.4-DPO](https://huggingface.co/NeverSleep/Noromaid-7B-0.4-DPO) Noromaid 7b v0.4 is a fully finetuned 7b Mistral trained on multiple RP dataset, modified by our own hand, and redone entirely from Alpaca to Chatml. The switch from Alpaca to Chatml and the addition of a new dataset from Aesir made it even better than 0.3. Noromaid 7b v0.4 DPO is the same as Noromaid 0.4 7b. Additionally, it was trained on top with 3 DPO datasets including the Intel DPO for reasoning, and some uncensoring. --- ## Spin-Off's _(standalone)_ ### Mixtral - [Noromaid-v0.1-mixtral-8x7b-v1](https://huggingface.co/NeverSleepHistorical/Noromaid-v0.1-mixtral-8x7b-v1) ~ a fine-tune of Mixtral 8x7b, trained on various RP datasets including LimaRP and Aesir, ToxicDPO without warning for decensoring and norobots, rewritten to use a modified Alpaca prompting to be on a par with ChatML or other conversational formats. This version contains 3x Alpaca modified datasets (for the RP one) and 2 shareGPT datasets. LimaRP token length, input, and output size got wiped out, and were separated into chunks of +8k context conversation (this is the maximum LimaRP can offer). - [Noromaid-v0.1-mixtral-8x7b-v2](https://huggingface.co/NeverSleepHistorical/Noromaid-v0.1-mixtral-8x7b-v2) ~ same base recipe as v1. On this v2, LimaRP got fixed further, Axolotl received a monkey patch to actually reformat shareGPT to our modified Alpaca prompting. - [Noromaid-v0.1-mixtral-8x7b-v3](https://huggingface.co/NeverSleep/Noromaid-v0.1-mixtral-8x7b-v3) ~ contains 5x Alpaca modified datasets. On this v3, all the datasets got trained on the "completion" method of Axolotl, with all the datasets being completely rewritten to be in Alpaca modified format. More than 2600 Wikipedia references got cleaned up of Norobots dataset. - [Noromaid-v0.1-mixtral-8x7b-Instruct-v3](https://huggingface.co/NeverSleep/Noromaid-v0.1-mixtral-8x7b-Instruct-v3) ~ made with a LoRA done on base Mixtral 8x7b. It's the same data than Noromaid-v0.1-mixtral-8x7b-v3, but applied on the Instruct model. Fine-tuning on base and applying on Instruct seem to give better result for our usage: RP/ERP. Stay tuned for more information! Noromaid-v0.1-mixtral-8x7b series info: v1 and v2 were each trained for 2 epochs, totaling to each 8 hours on Axolotl. v3, on the other hand, underwent 3 epochs and was trained for 12 hours, bringing the cumulative training time to 28 hours on a single A100 80GB GPU. ### Miqu ~ [HF Collection](https://huggingface.co/collections/NeverSleep/miqumaid-65c3d5e0fd15420346adc906) - [MiquMaid-v1-70B](https://huggingface.co/NeverSleep/MiquMaid-v1-70B) ~ Quick train to see if [miqu](https://huggingface.co/152334H/miqu-1-70b-sf) finetuned results in good models. - [MiquMaid-v2-70B-alpha-GGUF](https://huggingface.co/NeverSleepHistorical/MiquMaid-v2-70B-alpha-GGUF) (GGUF only) ~ trained on 1 epoch for 18h running on 2xA100 80GB. Trained on [miqu](https://huggingface.co/152334H/miqu-1-70b-sf). Deprecated! - [MiquMaid-v2-70B](https://huggingface.co/NeverSleep/MiquMaid-v2-70B) / [DPO](https://huggingface.co/NeverSleep/MiquMaid-v2-70B-DPO) ~ MiquMaid-v1-70B was our first 70B model, based on the leaked Mistral Medium model. V1 used Aesir datasets where V2 make the return of Norobots and some uncensoring data in it to make it even more unethical in RP.
V1 was already compliant to a lot of things, even on some prompt Mistral Medium would refuse 100%, because it is HIGHLY aligned. This V2 let you prompt even more unethical and unhinged RP. Not using any RP format show a rate of refusal really lower than Mistral Medium too!
On top of that, a DPO train, using the same data that was used in the OG Finetune for better performance made it even better, write better, and be even more uncensored. The model lose some points in benchmark, but the tradeback for really good RP and less repetition was worth it. - [MiquMaid-v2-2x70B](https://huggingface.co/NeverSleep/MiquMaid-v2-2x70B) / [DPO](https://huggingface.co/NeverSleep/MiquMaid-v2-2x70B-DPO) ~ really heavy, it's a 125B model made of MiquMaid-v2-70B and Mistral Medium base. Each MoE model have 2 expert active in them, so the idea between this was to have, on every token, 2x70B expert working together for more randomness and better precision. Since MiquMaid-v2 is here only for RP, it lose some IQ point, that's where Mistral Medium hit and make the prose a lot more better, and let the model be more logical. It's better than a frakenmerge of 2 Mistral medium because it's not 1:1 2x the same model.
The gem tho, is the DPO version. When the qLora for uncensoring was made on MiquMaid-v2 for uncensoring it, we got the idea to apply it to Mistral Medium base too. At first, the result wasn't really good and that was expected, since the qLora wasn't trained on Mistral Medium, but on MiquMaid-v2. BUT! When merged together (in a MoE for this example) the full potential of a double DPO shine.
We got astonish result on the worst quant ever: Q2_K. Even at Q2_K, MiquMaid-v2-2x70B-DPO showed godly performance in RP, following card, logic and smut. Only downside was the repetition of the formatting, but it was really usable. Unquanted is a gem, but we doubt anyone have the compute power to do that... _(based on 0.4)_ - [Noromaid-v0.4-Mixtral-Instruct-8x7b-Zloss](https://huggingface.co/NeverSleep/Noromaid-v0.4-Mixtral-Instruct-8x7b-Zloss) ~ It's Noromaid-v0.1-mixtral-8x7b-Instruct-v3 but retrained with ChatML, Zloss(Thanks charles), and some added datasets. - [FlatOrcamaid-13b-v0.2](https://huggingface.co/NeverSleep/FlatOrcamaid-13b-v0.2) ~ a merge between FlatOrca and Noromaid-13b-v0.2, see more info on the repo. - Important info to a 7b model: Currently not possible due to Orca-2 being a LLAMA 2 7b model and Noromaid-7b being a mistral model! - Important info to this model: This model is in NO way affiliated with [ddh0/OrcaMaid-13b](https://huggingface.co/ddh0/OrcaMaid-13b) or and \*maid models by [ddh0](https://huggingface.co/ddh0/OrcaMaid-13b), it was inspired by him tho. --- **Credits:** - Undi / Wrote parts of the blog post. ## Updates - **20.12.2023, 11:55:** Okay so, because some people asked about if the datasets will be released.. Yes they probably will, but not from me neither Undi. The maid part aka. Aesir will only be released after the official Aesir model from [MinervaAI](https://huggingface.co/MinervaAI) is released. - **21.12.2023, 19:56:** Added FlatOrcamaid-13b-v0.2 [ Released on: 20.12.2023 ]. Added Noromaid-7b-v0.2 [ Released on: 21.12.2023 ]. - **23.12.2023, 15:03:** Added Noromaid-v0.1-mixtral-8x7b [ Released on: 22.12.2023 ]. - **24.12.2023, 13:14:** Added FlatOrcamaid-13b-v0.2 disclaimer. Added Noromaid-v0.1-mixtral-8x7b-v2 [ Released on: 23.12.2023 ]. Added Noromaid-v0.1-mixtral-8x7b-v3 [ Released on: 24.12.2023 ]. - **26.12.2023, 00:18:** Added Noromaid-v0.1-mixtral-8x7b-Instruct-v3 [ Released on: 25.12.2023 ]. - **09.01.2024, 20:13:** Added NeverSleep Discord link. Added Noromaid 0.3 [ Released on: 05.01.2024 ]. - **12.01.2024, 19:13:** Added Noromaid 0.4 [ Released on: 11.01.2024 ]. Added Noromaid 0.4 DPO [ Released on: 11.01.2024 ]. Added credits. Added Noromaid-v0.4-Mixtral-Instruct-8x7b-Zloss [ Released on: 09.01.2024 ]. - **12.01.2024, 23:45:** Did some re-formatting. Fixed some skill issue mistakes on my side. - **05.02.2024, 20:14:** Added MiquMaid-v1-70B [ Released on: 31.01.2024 ]. Added MiquMaid-v2-70B-alpha-GGUF [ Released on: 04.02.2024 ]. - **07.02.2024, 20:20:** Added MiquMaid HF Collection. Added MiquMaid-v2-70B [ Released on: 07.02.2024 ]. Added MiquMaid-v2-70B-DPO [ Released on: 07.02.2024 ]. Added MiquMaid-v2-2x70B [ Released on: 07.02.2024 ]. Added MiquMaid-v2-2x70B-DPO [ Released on: 07.02.2024 ]. ### Echidna-13b and Nethena-13B, a collab with Undi (legacy) https://ikari.sillytilly.org/blog/legacy/echidna-13b-collab-with-undi.md Published: 2023-10-27T17:57:00.000Z Updated: 2023-10-30T23:50:00.000Z ### [Echidna-13b-v0.1](https://huggingface.co/NeverSleep/Echidna-13b-v0.1) Echidna-13b-v0.1 came out 21.10.2023, i wanted to merge it alone first, but had server issues. Thats when i asked Undi to help merge it, we reviewed my recipe and decided to upload it on his [NeverSleep](https://huggingface.co/NeverSleep) org and let it be a collab. ### [Echidna-13b-v0.2](https://huggingface.co/NeverSleep/Echidna-13b-v0.2) Echidna-13b-v0.2 came out 26.10.2023 (took 4~ hours of merging), it was kind of a "test" version of Echidna-13b-v0.3. Echidna-13b-v0.3 would follow soon after. ### [Echidna-13b-v0.3](https://huggingface.co/NeverSleep/Echidna-13b-v0.3) It took us some time to refine Echidna-13b-v0.2's recipe, we had multiple test models(private) which game way better results than Echidna-13b-v0.2. The newest version, Echidna-13b-v0.2 came out 27.10.2023 (took 5.30~ hours of merging). Repetition seems to be almost non-existant, which is really nice. Undi and me noticed that while trying to let the AI talk to itself(with impersonate). Coherency is outstanding, as is creativity. This model seems to not have 5 IQ like some other models, for example: a femboy has a dick, not a vagina... It's also really nice at describing things, like scenes. This is probably due to having so much medical data in here. Last but not least.. not a single "...shiver down X spine..." seen while testing! ## Updates - **31.10.2023, 01:50:** [Nethena-13B](https://huggingface.co/NeverSleep/Nethena-13B) ~ Tbh i cant really keep up with how much models we are posting on here kek.. Well anyways... we released a model called Nethena yesterday which is even better than echidna, this model combines every best model of Undi and me: Athena v3 + Nete + Echidna v0.3. We also made a [Nethena-20B](https://huggingface.co/NeverSleep/Nethena-20B) but no one really cares about that. #### Please enjoy! ### MinervaAI / Aesir (legacy) https://ikari.sillytilly.org/blog/legacy/minervaai-aesir.md Published: 2023-09-16T21:31:00.000Z Sooo.. hey.. me, Gryphepadar, Doctorshotgun, Lordgoonery, M1xx3l, Nruaif are in sort of(?) a little team making LLM models no. Currently we are working on a model + dataset called Aesir, we are almost 1 dataset manual cleaning down. Aesir is meant as a horny-ass RP model(ERP). You can get more info on our [HuggingFace](https://huggingface.co/MinervaAI). ### Wanna check out my resources page? (legacy) https://ikari.sillytilly.org/blog/legacy/resources-page.md Published: 2023-08-14T21:17:00.000Z Updated: 2023-08-14T22:49:00.000Z So i added a little section called "Resources", if you want you can check it out if you wanna look for new stuff programming/ML/LLM/SD or if you are starting out with that stuff. You can check it out [here](/resources/) or you can just click the "resources" button in the Navbar. One thing i still need to do tho, is making this website better for mobile, as it looks shit currently(i think so at least). ## Updates - **15.08.2023, 00:49:** I also added a projects page, where i will post future projects. ### Website is finally "done" (legacy) https://ikari.sillytilly.org/blog/legacy/website-is-finally-done.md Published: 2023-08-13T20:29:00.000Z Updated: 2023-08-13T22:39:00.000Z Hey, i finally made the website here have content, and if you are asking "ikari, why tf are you posting so many blog posts, aren't they suposed to be kind of 'special'?" To that i answer "Shut up, i need to fill this page so it doesn't look empty!" Anyway.. it now looks awesome! And i removed all the lorem impsum shit, + I added, if you open a post in "full screen" it temporarily removes the sidebar for better visibility of the post. ## Updates - **14.08.2023, 00:39:** Added a nice looking loading screen, to hide the js magic, and fixed a couple of things, + quality of life stuff. Also added a vote thingy: if you like a post you can vote it up (you can for some reason press it multiple times, so if you REALLY like a post you could vote it to like 200 or smth). ### So i added a couple of cool things (legacy) https://ikari.sillytilly.org/blog/legacy/a-couple-of-cool-things.md Published: 2023-08-12T21:05:00.000Z I added a new url arg called "bo", when set to true combined with set blogid arg, it opens it in "fullscreen" mode. I also added a little button(Open post), which does the same. When you click share it automatically sets the blogid arg, + it sets "bo" arg to true. ### Hey, this is a little test for the new blog page (legacy) https://ikari.sillytilly.org/blog/legacy/new-blog-page-test.md Published: 2023-08-12T02:15:00.000Z Updated: 2023-08-12T02:33:52.000Z I will document my jurney as a dev here. I dont think anyone is even reading this but yea, i dont really care xD If you wanna for some reason share this or other posts, i added that thing below