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

Trigger phrase: Japanese color film photograph

Trained mostly on scans of late 80s, early 90s Japanese softcore adult/gravure magazines/photobooks printed on high quality paper, with a small number of newer or older images. Over 200 images total, mostly even mix of nude, clothed, partial clothed women with era appropriate hair/makeup/fashion, and mostly (85-90%) of Japanese women.

The LoRA imparts an analog photography/print media look with somewhat warm and grainy colors. Text was mostly left untouched in the dataset, but should not appear unprompted.

---

Generation advice:

Weight: 0.80 to 1.0 should be fine. I went with 0.88 for all the sample images and it just worked.

Hires. fix and ADetailer should not be needed! I normally always i2i upscale my images, Flux or not, but for the samples only the cover image is inpainted and upscaled. Rest are straight txt2img results, though some use a variation seed.

Should do fine with standard 1MP resolutions. I mostly generated 768px smaller side since that's heavily in the dataset, but it probably didn't matter.

Note: All sample images were generated with the original save as float model. The uploaded resized fp8 version should be practically identical, though I didn't re-test every single image.

See sample images for prompting examples. LoRA was trained on descriptive formal JoyCaption generated captions, and being wordy works for prompting it too. Specifying a time period ("mid-to-late 80s, late 80s, late 80s or early 90s, early 90s" appear in the captions) might help.

Tested with fp8 version of Flux-dev, and Flux.1-Dev BNB NF4 v2. Should work with Krea.

All sample images are Euler Beta distilled CFG 3. DPM++ 2M Simple can improve anatomy and even general coherence, but doesn't seem to engage the style as much.

Not tested with other LoRAs. Should do fine by itself for the most part as anatomy is pretty good, and it does mostly know what breasts and nipples look like, but could use a little help with nipples. Also, no image in the dataset shows actual genitals, censored or not, just pubic hair, not that base Flux knows what genitals are anyway so it's just as well.