Misha Movernco - Venus Project: Climax - Pony

Hello! I wasn't originally planning on sharing this LoRA as I feel there is still much for me to learn about making a proper one. A user requested I make an earlier version of it public, but it wasn't very consistent and had some major flaws at times. I retrained it and now it's more consistent with errors occurring less often, so I feel that it's in a state where I won't feel bad sharing it. You can see some of the things you may run into while generating in the showcase grids and the images they are made from.

The grids and individual grid images were generated randomly without cherry-picking or editing. They have visual errors. I did this to hopefully show what you can expect from using this LoRA. Each of those images includes the metadata. I generated using the older Forge WebUI (f0.0.17v1.8.0rc-latest-278-gbfee03d8, prior to the Flux update).

The trigger is "mmovernco". It is intended for use at a strength of "1", but lower may yield better results in your specific use case. For her default business outfit seen in the anime, try:

• "mmovernco, black suit, pearl necklace, yellow shirt, collared shirt, brown hair, cleavage, cyan eyes, black hair tubes, large breasts, semi-rimless eyewear, short hair, sidelocks, red gemstone, black pants"

For just her physical appearance and hair (to use with other outfits):

• "mmovernco, brown hair, cleavage, cyan eyes, black hair tubes, large breasts, semi-rimless eyewear, short hair, sidelocks"

The above prompts seem to work well despite being slightly different than I used when training. Removing "1girl" from the prompt seemed to look better most of the time as well. I don't typically use "score_9" in my positive for the same reason. I do use "score_8_up,score_7_up,score_6_up,score_5_up,score_4_up, " in the positive prompt and "abs, source_anime, score_6_up,score_5_up,score_4_up,score_3" in the negative. The "score" tags have a noticeable effect, but I am never sure I am using them correctly. "abs" is in the negative because I used two AI generated pics of a fairly muscular version of her in the training data and that came through pretty strong with my preferred model. Your own results may vary.

I trained on "ponyDiffusionV6XL_v6StartWithThisOne.safetensors" hash: 67AB2FD8EC (all hashes listed in this description are AUTOV2 hashes) but found that I got the most consistency and accuracy with "Autism" models. My best results were with these settings:

• Steps: 30, Sampler: DPM++ 2M Turbo or DPM++ 2M SGMUniform, CFG scale: 4.3, Model hash: 2bcedd17aa, Model: autismmixSDXL_autismmixLightning, VAE hash: 235745af8d, VAE: fixFP16ErrorsSDXLLowerMemoryUse_v10.safetensors, Lora hashes: "misha_11: 774cbf4e7323"

The steps, model, sampler, and CFG are not typos. I realize that the model is a "lightning" model, but of the publicly available models, it's what worked best for me. I do not know nearly enough about how this tech works to understand why that one worked better than the others I tested, but I probably did something wrong. You may have better luck with something else.

I also use the following WebUI settings:

"Pad prompt/negative prompt (improves performance when prompt and negative prompt have different lengths; changes seeds)" is checked and my random number generator is set to "GPU". I do not know if that affects results for others or just myself.

Some additional notes:

• I took some liberties with the necklace. It was just a few circles in the anime that (to me) looked like pearls with a gemstone. The results don't always get it right and "red gemstone" frequently attempts to put gems in other places than just the necklace. If you prefer a different style of red gemstone necklace, generating with just "necklace"(instead of "pearl necklace") and "red gemstone" usually makes a thin chain with a decently large red jewel attached to it.

• Some angles and image dimensions are more problematic than others. If an image isn't cooperating, try adjusting the dimensions. It may drastically improve the result.

• I originally tried for "rimless eyewear" since I believe that is closest to what the anime intended, but I could not consistently get rimless glasses and eventually just went with "semi-rimless eyewear" instead.

• One of the most common problems is the colors. I probably overbaked it. The shirt under the suit will sometimes be the wrong color. You can also see that some colors occur more frequently by default. Prompting for specific clothing colors helps alleviate the issue. The glasses will often appear with color rather than the simple metal seen in the anime. Hair tubes can also appear in other colors/styles. Many of these issues occur less frequently with the model and settings listed above.

• The LoRA is capable of both SFW and NSFW.

This is my first upload to the site. I still don't think I'm very good at making these, but I hope that some of you can find some enjoyment from this!