This is my first LoRA.

I haven't been using enough parentheses in my prompts! The LoRA may be much more flexible than I thought.

Usage (v1.1)

No embedding needed for this version.

Trained on Pony Realism v2.2 Main + VAE, which does not support DPM++ 2M Karras - see its page for the recommended ones like DPM++ 2M SDE Karras, DPM2 a and Euler a. Doesn't work with Pony Diffusion V6 XL. Not tested with other models.

Prompting is mostly similar (see the images for 1.0 for ideas), but the word "inflatable" may trigger colorful blobs instead of brown equines (this may be a sign of overtraining or CFG is too high). This makes it harder to generate inflatable horses and whatever was not trained enough.

Thanks to dataset balancing, "jets medium zebra" and "jets small zebra" can be generated with some effort or with minimal prompting.

Usage (v1.0)

Trained on Pony Diffusion V6 XL.

Important

This version requires this embedding file (may have a slightly higher rating): https://civarchive.com/models/1041364?modelVersionId=1168345 (put it into the folder for embeddings and/or refer to it in the prompt as described). The training workflow I used recommended enabling embeddings, so I didn't use the word "inflatable" anywhere to save space for other tags.

Include something like this in the positive prompt when running on ComfyUI (the part after the ":" is the file name, so edit it if needed - pay attention to the number of "embedding_"):

embedding:embedding_embedding_embedding_zeblora-20241214-8-object_sdxl_embeddings

Keywords in the same tag as that string seem to relate closer to the LoRA. If the embedding cannot be found by that name, ComfyUI will print a message in the terminal.

Example prompt that works on Civitai:

zebra embedding:urn:air:sdxl:embedding:civitai:1041364@1168345, photo

The "photo" part may not always be necessary, but I couldn't get anything inflatable right without it.

Training

Based on Pony Diffusion V6 XL, it was trained on about 327 photos of inflatable equines and other toys without humans, but by default it will generate a zebra most similar to the 6ft Jet Creations zebras that were most popular.

About 3+3 images in the v1.0 set contained NSFW details, but I could only retrieve poses (stacking).

Tags tried:

Top tags that work with the embedding (with the number of photos, although a couple of them had too many tags, so some were skipped):

• zebra (320)

• b/w coat (302)

• jets big zebra (210)

• indoors (149)

• outdoors (82)

• horse (78)

• lying down (44)

• tiger (36)

• side view (22)

• zecora (15)

Might do something, but not necessarily what I meant by adding them:

• inflated enough (286)

• feral (261)

• fully inflated (245) - implies "inflated enough"

• jet creations (154)

• jets llhorse (41)

• iw (38)

• iw horse (36) - may or may not generate an Inflatable World horse

Other interesting tags:

• deer (30) and jets deer (28) generate an inflatable deer, but not the kind that was in the photos.

• jets medium zebra (77) - does nothing special.

• vbs horse (60) - probably same as horse

• pinto (60)

• orange coat (47), brown coat (21)

• standing (25)

• butt (22) - artifacts

• plot (20) - does not work, artifacts

Redundant:

• photo (322) - add this if you are getting non-inflatable 2D images

• nonpop (318)

Training

Base model: Pony Diffusion V6 XL

Dataset: about 327 photos of inflatable equines and other toys without humans, found by Jet Creations zebras, most of which were the second-generation 6ft ones. Inflatable World zebras were not searched by, and the few that were found must have blended with them.

edenartlab/sd-lora-trainer

After choosing 512x512, it automatically resized and cropped everything to 576x447 or 447x576. I think it said it cropped to the center of mass.

It had no mention of the balancing features available in kohya_ss.

It recommended using an embedding, which is uncommon on Civitai.

You are supposed to choose your own prompts for checkpoint test images, and I did not. For my previous attempts, I had increased the number of test images, so it used the captions as prompts, but this time it used predefined prompts, most of which were irrelevant and not helpful.

Copyright

The permissions I selected were based on the following considerations.

The LoRA was based on Pony Diffusion V6 XL, which has a license that might apply.

The most generatable toys were sold by Jet Creations, Inflatable World, Hongyi. Toys by Horseplay Toys, G&G and PuffyPaws may have influence. They may have other authors and copyright holders.

The source photos were found on the Internet. The output is very generic, but if you recognize the generic setting of your photos and want attribution, feel free to ask.