BudgetFurry

This is an SD1.5 budget model focused on anthro and furry subjects, but can do other stuff too. It is intended for offline generation on slow hardware. Use a higher-quality model for online generation, like BB95 Furry Mix v14. This is a modification of that model.

Examples

The example images are raw generations, upscaled twice, without edits. The sole exception is the removal of fake watermarks, which sometimes show up. That's done with a single stroke of the Smart Patch tool in Krita. All images are generated in ComfyUI with the exact same seed, in batches of two images.

There's an example workflow to get you started in ComfyUI. I'll add more creation guides in the future.

Workflow

This model doesn't produce high quality images, but it's fast. It needs 8 steps, and doesn't use CFG. That means with the right software (I use ComfyUI) it only takes one sample per step instead of two. So compared to a typical 20-step SD1.5 image generation this is roughly five times as fast.

Base Size: 512x768 (portrait), 768x512 (landscape), or 512x512 (square). You can try other sizes too but these I have extensively tested.

Steps: 8. No more no less.

CFG: 1. No negative prompt. Deal with it or use another model.

Scheduler: SGM Uniform. Anything else is a trade-off that we cannot afford to make with so few steps. If your software doesn't allow separate scheduler selection at least avoid the Karras variants.

Sampler: A fast one, anything else is defeating the purpose of using this model.

Straightforward Samplers

Euler (soft), DPM++ 2M (sharp), and LMS (rough). These usually produce similar results, each with their own flavor.

Creative Samplers

Euler Ancestral (soft) and DPM++ 2M SDE (sharp). These typically each go their own way, though sometimes all samplers produce fairly similar results. ComfyUI has two SDE versions, one using the CPU and another the GPU for noise. There is no significant performance difference, they just produce different results.

Upscaling

I generate at a base size until I get something worth keeping. These images are unrefined and can be considered a sketch. Then I upscale x1.5 using either a pixel-space upscale with Lanczos filtering at 0.25 denoise (sticks close to the base), or a latent-space upscale with either Bicubic or Nearest-Exact filtering at 0.5 denoise (higher quality but deviates from the base). It varies which approach is best, so I often try two or all three. Use 4 steps for all variants.

If the upscaled image is OK you can stop there, but I always do a second x1.5 Lanczos upscale so the final result is x2.25. You could go higher but this is a good trade-off between quality and speed.

For simplicity I usually keep the same sampler for both the base and upscale phases, but you can switch them too. For example, upscale a nice Euler base with DPM++ 2M to make it sharper.

Editing

I never edit the initial generation, just keep iterating until I get something nice. Editing and touch-up is done on the x1.5 upscale. Rough edits then get smoothed out by the final upscale. Minor touch-up can be done on the final upscale.

A useful trick is to manually mix multiple upscale results of the same base, combining the best parts of them all. This can also be done with different the results of samplers, if their results are similar enough.

NSFW?

This model can generate nudity and can do so unprompted. Prompt for clothes to avoid that. While you could try to generate explicit sexual acts, it's not good at it. All but the simplest stuff will produce a mess. Don't bother and use another model for that stuff.

Why?

I made this model because I like to create funny, silly, cute, and sometimes sexy images, without having to wait long before I get some results on modest hardware. I'm comfortable with editing images manually, so prefer generation speed over quality. I experimented with models and LoRAs until I settled on a mix I liked, which I solidified into this model.

Recipe

This is BB95 Furry Mix v14 merged with Hyper, DPO, and SPO at various strengths after a lot of experimenting. DPO and SPO balance each other and enhance the hyper quality.