This Lora is amazing at zero-shot generation!

Trained on 297 images I manually processed (and stole from the internet respectfully. I mean I did it by hand so that's a little better right?). This probably the best 2D Murder Drones Lora out there. Let alone one based on Chroma1-HD. It's extremely flexible and can generate a broad range of sfw and nsfw material. Using feels like it was trained on a lot more than just a few hundred images.

v5-fullHD

This version has been retrained on the same dataset, but in full 1024p and in higher rank. It's much better than the previous versions in many ways.

The model works best with a base resolution of 1024p, instead of the previous 512p.

Additionaly it doesn't really need Reakaaka's Enhancer Lora anymore.

v5-fullHD: ComfyUI + 1024p + "DPM++ 2M SDE" + Beta Scheduler + CFG 3.5 + 30 steps

v3-HD: ComfyUI + first 512p -> upscale -> then 1024p + "DPM++ 2M SDE" + Beta Scheduler + CFG 3.5 + 30 steps

Works best with:

• Reakaaka's Enhancer Lora: https://civarchive.com/models/2014953/reakaakas-enhancer-chroma-hd?modelVersionId=2280444

• In 512p and 1024p

  • Optimally in a 2-stage-process (generate in 512p then upscale and do a second pass at 1024p) for speed and composition quality.

  • 1024p on its own is a little rough (but can still give great results, just more artifacts)

• ComfyUI

• Use DPM++ 2M SDE! (It's so much better than Euler! Sadly Civitai's online generator uses Euler as default for Chroma.)

  • Euler tends to make images too blurry and requires "blurry" in the negative prompt.

• CFG of 3.5 should be fine. Optimally you can use Chroma1-Flash-HD or Flash Lora with CFG of 1.

Prompting:

• Works best if you start the prompt with the character tag: SD-V, SD-J, SD-N, Uzidoorman, Cyn

• If a character isn't generated correctly, type more describing traits: SD-N -> SD-N with short hair, wearing a black coat

  • There can be mixups with a too short prompt.

• The Lora likes to forget the tail. Luckily you can remind it by typing: tail

• Natural language prompts work best, but can be combined with tag based prompting

• Result is badly drawn and whatever? Chroma generally likes to just have more text. Helps most of the time.

  • Either way, you should try Reakaaka's Lora

• Sometimes the negative prompt can mess things up. Keep it minimal!

Bias:

• The Lora is heavily biased towards generating V, J and Uzi. N and Cyn need a few more describing words.

• Most APIs I used to get images are on the more nsfw side of things so... write "wearing clothes" to hide that bias.