<h2 id="description">Description</h2>Experimental models. Most are not exactly full-fledged fit for all use cases, and potentially will only work with certain keywords. Please don't use commercially but feel free to remix. Negative embeddings not necessary for quality but might be helpful.Heavily modified <a target="_blank" rel="ugc" href="https://civarchive.com/models/14734/store-bought-gyoza-mix">Gyozamix </a>recipe. Here's a quick rundown of the differences between each model:<ul><li>MIX-GEM-T1: This one is targeting a kind of oil canvas-like texture. It's a very primitive LoRA tuning, just mixing two artstyle trained LoRAs together at a dominant-recessive ratio. The framework it's built on is mostly ad-hoc, as it's the result of some random mixing with surprisingly good results. Decently responsive to prompts.</li><li>MIX-GEM-T2: (Also called EternalWinter). This one is built on the earlier oil canvas fine-tuning, but with additional work put in to affect a very delicate sort of look. There hints of the mangaka of Galko, Suzuki Ken'ya's style in there. As a result it has very delicate features around the eyes and especially the lashes.</li><li>MIX-GEM-T2A: (Un-uploaded) A further refinement of T2. Overall it comes with more details and likes to do even more contrast.</li><li>MIX-GEM-T3: Due to some quirks in the merging process and a failure to account for adjustment in merging materials, this one is stylized with a heavy cartoon influence. It loses some of the delicateness from T2 but also gains more exaggerated features. It's an interesting line to study down towards, though it's not really my preference, so I did not continue down this avenue of investigation.</li><li>MIX-GEM-T2_2: A more comprehensive rework of T2. Essentially a MIX-GEM-T3 recipe with the imbalance that led to heavy cartoon styling in T3 rectified. It is further enhanced with a CLIP-swap for more prompt-obedience.</li></ul>I like to use loras from <a target="_blank" rel="ugc" href="https://huggingface.co/ashen-sensored/lora-isolation-collection">here </a>to modify the output. Character, outfit, and pose loras are fine. As this model is stylized already I don't really see the point in using a style lora. It's not intended to be versatile, use Based66, AuroraOne or Amedira for that.No HuggingFace repository for this one, recipe (if I bother to provide it) is in the sidebar.<h3 id="sdxl-update-4152024">SDXL Update 4/15/2024</h3>We're back with SDXL, lol. Sorry to @penterested, i lied, lol. Using Pony for now since my favorite artists haven't been trained on AnimageXL yet:<ul><li>MIX-GEM-WXL-D2nai: Uses <a target="_blank" rel="ugc" href="https://civarchive.com/models/315596?modelVersionId=354017">Pony Diffusion for Anime </a>as a base. Worldly_XL is based in at a base level to compensate for Pony's deficiency in backgrounds. It's pretty compatible with LoRAs. I was considering releasing it as a T(aste) version, but novelai is probably the only sdxl model that can survive on the inherent styles within the model and you should just use LoRA to compensate for local.<ul><li>MIX-GEM-WXL-D2naiK: A variant of D2nai with a style lora already baked in. Still compatible with other style loras.</li></ul></li><li>MIX-GEM-WXL-ZA1: Also known as <a target="_blank" rel="ugc" href="https://civarchive.com/models/498178/chromeshinexl">ChromeshineXL</a>. This follows in the wake of D2naiK of imparting an inherent style. The primary advantage compared to D2nai is that it has upgraded lighting and better concept recognition for robot girls and monster girls (particularly centaurs). It moves further away from D2nai's realistic bias, the amount of prompt keywords required to remove realism is further reducing.</li><li>MIX-GEM-LXS-QromEW: This version moves away from the PD for Anime base. I'm not sure yet if it's a one-off case or if it will be the case going forward. It's basically built to serve one single style and it does it fairly decently. It uses <a target="_blank" rel="ugc" href="https://civarchive.com/models/529439/chromelightxl">ChromelightXL</a>, which is extracted from MIX-GEM-WXL-ZA1, but is otherwise not connected to the rest of the MIX-GEM-XL heritage. Probably a dead end evolution tree here but I get the impression that <a target="_blank" rel="ugc" href="https://civarchive.com/models/494290/ls-equos-xl">LS Eqos</a> is a fairly powerful and versatile base so there might be some parallel evolution. This model should be completely unresponsive to style loras.</li><li>MIX-GEM-LXS-ZA1: (Un-uploaded) An attempt at transposing ChromeShineXL's unique qualia onto LS Equos. Didn't really work out all that well. LS Equos + <a target="_blank" rel="ugc" href="https://civarchive.com/models/529439/chromelightxl">ChromeLightXL </a>seemed to have better results. Compared to ChromeShineXL there were less watermarks, the backgrounds were better, and the posing seemed to be more dynamic but it lost understanding a lot of a lot of character2character interactions and took major downgrades in lighting and tones.<ul><li>MIX-GEM-XLS-ZA1.trainD: Ended up training the initial differences from D2nai onto LS Equos and merging to finish the final touches. This is <a target="_blank" rel="ugc" href="https://civarchive.com/models/498178?modelVersionId=605702">ChromeshineXL_MP</a>. The actual model page has more detail on the comparison but basically it's way easier to use and less finicky with LoRAs at the cost of some composition sense.</li></ul></li><li>Adam8wit-QEgooW: The sidebar is way more comprehensive than I intend to be in this tab, but essentially it's a heavily redone merge to try to get some use out of the <a target="_blank" rel="ugc" href="https://civarchive.com/models/316075?modelVersionId=354534">PonyXL T2_2 LyCORIS</a>. It is extremely clean, is great at skintones and lighting, and has fairly decent backgrounds (for Pony, anyways) but has a tendency to stray towards the 2.5D side of things. My recommendation is to use it with character LoRAs when possible to mitigate excessive 2.5D-ness. Unless you like that sort of thing, of course. In that case, go wild. Combine with <a target="_blank" rel="ugc" href="https://civarchive.com/models/529439/chromelightxl">ChromelightXL</a> to get the milf-faces back again.</li></ul><h3 id="nai-xl-update-1112025">NAI-XL Update 1/11/2025</h3>Personally I don't believe in Illustrious-based style mixes because the text encoder on Illustrious is not cooked as bad it is on Pony. However, I'm experimenting with some Itercomp and Cyberfix stuff right now because I merge them into the base model to improve compositional sense and general anatomy so i figured I might as well see how it works.So far my conclusion is that it's actually pretty impressive for base models but they suck for style models.<ul><li>XLVpred06: Uses <a target="_blank" rel="ugc" href="https://civarchive.com/models/940235/lhc">LHC (Large Heap o' Chuubas)</a><a target="_blank" rel="ugc" href="https://civarchive.com/models/315596?modelVersionId=354017"> </a>as a base, which in turn uses NAI-XL vpred06 as its own base. In all honesty I only used LHC over Nai-XL is because I deleted vpred06 to make room on my disc, but it has been pretty good and has offered an expansion of character knowledge over base NAI-XL. This model is closer to the SD1.5 versions than ever before and also is built on a way better SDXL base compared to the frankly quite flawed PonyXL.<ul><li>XLVpred06DOUBLEcompfixed: For the record, I don't like the style as much as XLVpred06. But this is the primary reason I was poking around with this merge. Primarily the double-compfixed version has undergone two processes compared to XLVpred06. It has been perpendicularly merged with Itercomp in order to improve style and composition sense, and then perpendicular merged with Cyberrealistic and Cyberillustrious to improve anatomy. The Itercompfix version is merged with both the Cyberrealistic and and the Cyberillustrious-fixed versions and then both are merged with each other to get the double-compfixed version. Overall it gains more than it looses, particularly in regards to hands, however it dilutes the style. I'm planning to try the same method to Noob itself, where with it being a base model the style dilution won't matter as much.</li></ul></li><li>XLBD1: Uses <a target="_blank" rel="ugc" href="https://civarchive.com/models/623891?modelVersionId=1143094">Better Days. Illustrious v0.1</a> as a base, specifically a remerged version of it called <code>betterDaysIllustriousXL_V01ItercompPerpCyberillustfixPerpWeightedSum05</code> utilizing an abbreviated version of the XLVpredDUBLEcompfixed method. If you're really interested in how it performs without the lora merge, you can find it <a target="_blank" rel="ugc" href="https://huggingface.co/Jemnite/experiments/blob/main/Better%20Days/betterDaysIllustriousXL_V01ItercompPerpCyberillustfixPerpWeightedSum05.safetensors">here </a>or <a target="_blank" rel="ugc" href="https://huggingface.co/LMFResearchSociety/Tiantianquan/blob/main/SDXL/Noob/betterDaysIllustriousXL_V01ItercompPerpCyberillustfixPerpWeightedSum05.safetensors">here</a>. It is basically Better Days with some lighting improvements and much less Sepia. I posted some x/y comparison pics <a target="_blank" rel="ugc" href="https://civarchive.com/posts/12659123">here</a>. This version is eps because of the eps base.</li><li>XL-ŞİHKKLZemb(MFCG): After fucking around with component swap in ŞİH, we have this abomination. The primary goals here were to extract character knowledge from models with a later cutoff date for training data than base NoobAI (failed) and to reorient composition bias (partial success). Generally speaking this is not a good model but the damage done to Layer 2 and 3 created a semi-interesting color palate and fritzing effect that I thought was worthwhile to add to the model repo here. It's a strange model that only does a few specific things well in a way that I can't replicate with a better model.<ul><li>XL-ŞİHKKLZemb(dwnsty): Simple swap of the 0th layer text encoder model. If you want to know what changed, I used the one from NovaAshMix instead of MFCG's Soft Serve (hence the change from MFCG -&gt; dwnsty). I like the composition better and feel that it gives cleaner outputs, but in general this shares all the weaknesses of the previous model.</li></ul></li></ul>

MyMIX-G/Jem

Swapped out layer 1 and text encoder. Model ends up being slightly better at composition. Additionally the less fucked up TE means the initial gen is less fucked up, relying less on secondary denoise cleaning. The soft style is somewhat tied to the noise, can't seem to get rid of it without shifting styles into something with cleaner lines.

Files with SHA256: f6b1c20436a52ef9245c1e9b33ff050409ba7737870fdf3c7d2475004248e354