Retro 80s-90s Japanese Softcore Adult Photo Style 【JP 雑誌 / 写真集】

28.07: much belated note about the example images: I used the DDIM CFG++ sampler for many of them, because I liked the results. This worked fine on A1111, but I have since mostly moved on to Forge where DDIM CFG++ outputs garbage for me. I don't know why, but you can just use DPM++ 2M SDE for example, for similar enough results.

---

Trained mostly on late 80s, early 90s Japanese softcore adult magazine/photobook scans, with a small number of newer or older images. Over 150 images total.

The LoRA imparts an analog photography/print media look with somewhat warm and grainy colors. It probably cannot reproduce magazine pages with pseudo-text and such, as most of that was cleaned out of the dataset. It's somewhat horny as more than half the dataset was full or partial nudes.

Generation advice:

Weight: 0.80 - 1 works fine. You can try other values depending on prompting, checkpoint, use with other LoRAs, and what you are trying to do.

Hires. fix and ADetailer (mediapipe_face_full mainly) are both highly recommended, but skipping hires is plausible.

Clip Skip: I use both 1 & 2, and noticed that 1 tends to delete almost all pubic hair.

Trigger: jpg4ls photo, woman

Useful tokens for controlling the output: asian, brown eyes, black/brown hair, short/long hair, bangs, large/medium/small breasts, nude, female pubic hair, etc.

Black or very dark brown eyes were not tagged and are baked into the LoRA, which Adetailer will sometimes want to "correct" to be brown.

"Bangs" corresponds to 80s-90s style feathered/wispy bangs which are heavily represented in the dataset. You may have to prompt them in the negative if you don't want to get them.

Tokens like vintage, 80s, 90s, 1980s style, analog, etc. are not in the dataset but may help, depending on the checkpoint.

See sample images for some starter prompts. They are all straight txt2img + hires + adetailer results, generated locally on A1111. No other LoRAs were used.

Dataset, training settings and tagging could all use a revision so an update is possible at some point, but I want to release an SDXL version first.

If there's anybody still checking out SD1 LoRAs, any feedback is appreciated and please share your results!