I'm going to be restarting this project with hopefully better organization and methodology!! For now, all my development versions since the first public release have been made available. They're not fantastic, but hopefully one of them does what you need!
UPDATE: I'm planning to restart work on this at some point, but in the meantime, the excellent Inflation by Budace exists! It's trained on FluffyRock instead of Fluffusion, which the new incarnation of InflateBot will also use.
InflateBot is a joke that went way too far; originally an attempt to fine-tune GPT2 on inflation ???????, its latest shape takes the form of a LoRA based on the excellent Fluffusion model.
The goal of InflateBot is to provide multi-target inflation and full-body (spherical) inflation that isn't just a featureless blob. Its training data is multi-concept, with folders for the following tags:
"air inflation"
"blueberry inflation"
"floating" (where it coincides with an inflation tag)
"hyper inflation"
"inflation fetish"
"spherical inflation"
The issues targeted by InflateBot are:
Over-focusing on belly inflation (ignoring body/spherical/hyper tags)
Creating blank spheres with no character at all
Creating a sphere that is detached from the character, who may be relegated elsewhere in the image
Not generating arms or legs.
In the past, putting "balloon fetish" in the negative prompt helped alleviate some of the issues, but the goal is for that to no longer be necessary. The first demo image below uses this technique, the second does not (Private Version 2 "Oxide" used for these demos)
While InflateBot isn't done yet, I feel like it's far enough along that I'm willing to share some snapshots and see what you guys can do with it! (Note that InflateBot utilizes LyCORIS training methods, which may or may not be relevant depending on your UI. If you don't know what that means you probably don't need to worry about it.)
Description
THE OXIDE SERIES USES THE LORA EXTENSION, NOT KOHYA/LOHA LIKE THE OTHERS, POSSIBLY ACCOUNTING FOR INCONSISTENCIES IN MY TESTING. THIS IS MY FAULT.
Hey, long time no update!! I've been busy moving house, and as such, no work has been done on new versions since the last update on the Furry Diffusion Discord server. However, while experimenting with these old versions, I found that this one, which was released to FD but not here, has been giving much better results on full-body inflation than later iterations that were (I believe) unreleased.
Other things that help are putting the trigger tags early in the prompt (as it's expected that the character will take up a lot of space, it's best to block out the shapes for them early) and putting "limbs missing, limbs hidden" in the negative prompt. (This was supposed to be trained into the later versions, I'm not sure if that dataset change made it into this version, but it seems to make a difference regardless.) Dioxide was trained on Fluffusion e12, but works great on e20!!
The plan for the next release is to basically start from the ground up, with a more consistent dataset and more thought put into its structure; however, these "throw things at the wall and see what sticks" experiments are promising.
(does the "steps" field mean "steps per epoch" or "steps total"? do i multiply it by the batch size? civitai is weird lol anyways the step count represents 4 epochs of 994 batches each at batch size 6)
FAQ
Details
Files
Available On (7 platforms)
Same model published on other platforms. May have additional downloads or version variants.
