InflationMix (InflationXL Merged LoRA)

NOT A CHECKPOINT!!!

See below for FAQ

A merge of the excellent InflationXL LoRAs by Tigerpillow (shared with permission):

Inflation XL

Male Inflation XL

Intended to allow for a wider range of body configurations (e.g. women with ???????, boys with ????) in a steerable fashion. Made with SuperMerger.

Known issues:

• It can still do rounds but it takes some extra fiddling.

• "Blimp" tags have been severely attenuated and basically do nothing without an accompanying "hyper" tag. (NOTE: With further testing, the dimension rebuild that was the last step for the current version ended up helping this a lot!!)

• Total sizes are somewhat constrained compared to the sources. This is because the larger sizes are typically only achieved with higher weights that tend to introduce side-effects, and I tried to optimize for quality. Again, more extreme sizes are still doable, but aren't so much the goal. "Over-weighting" may help, but may also introduce artifacts.

• With male characters especially, any "inflation" tag tends to cause bellies to show up where you don't intend. Putting "obese" and/or "weight gain" in the negative prompt can temper this. This behavior is observed without any LoRA at all, and is tempered by using the base InflationXL LoRA; it seems the male LoRA reintroduces the problem.

FAQ

Q: Which version do I use?

A: They're ordered by what I recommend. The block-weight merges came out way better than I anticipated. Failing that, try the standard merge. The "no-base" models are rather unstable.

Q: Why wouldn't I just use both LoRAs at the same time?

A: You absolutely could! But a merge has its advantages. Using the standard merge will give you a decent result most of the time. The block-weight merge is designed to help alleviate a problem with using both Inflation XL LoRAs together, where the belly expansion data from Male Inflation will reintroduce the tendency of Pony to produce bellies where it isn't intended. Also, if you use the non-rebuilt LoRAs, you save on VRAM. (I'll have those up for the rest of the published models soon, but posting to Civit is kind of annoying TBH.)

Q: What does r128 mean?

A: It means that I had "Rebuild Dimensions" set to 128 in SuperMerger. Essentially, while merging the models, a larger pool of weights was allocated than for either of the base LoRAs. The upshot of this is that background detail is better preserved, and concepts from both tend to sit more seamlessly with each other. However, this also means that tendencies of the base model that may be problematic (see above) will have more chance to leak. Overall, most of these merges exhibit better detail preservation and less style swing than either base LoRA at full strength, which is a pretty astounding achievement in my book.

Q: Why wouldn't I just do this myself?

A: Be my guest!! You absolutely should do that. Here's the recipes I used. SuperMerger was used for the merge, LoRA Block Weight (same developer) was used to test recipes prior to merging.

I'm standing on the shoulders of giants. Tigerpillow basically ate my lunch but you know what he can have it. Go support him on Ko-Fi.

Use it to make friends, not wealth. :)