This is a LoRA I created for a commission of 700 images since none of the

other ones available were high enough quality of her DC Animated Movie

Universe appearance.

## Links

- Patreon: https://www.patreon.com/c/Anzu777

- Full post of 700 images: https://www.patreon.com/posts/156920862

- Pixiv: https://www.pixiv.net/en/users/125260991

##### Use these tags if you want to get the same images as in the examples above with Weight:0.9 CFG scale at 2 and face detailer (do not include cel_shading, flat_color)

r4ven_dc, short_hair, red_gemstone, pale_skin, red gemstone on forehead, forehead,

black_leather_jacket, cropped_jacket, open_jacket, short_dress, blue dress, dark blue skirt,

necklace, jewelry

####

## Trigger word

r4ven_dc

## Quick prompts — pick blocks and combine

Always include (character + style locks):

r4ven_dc, 1girl, solo, short_hair, bob_cut, cel_shading, flat_color, cartoon, looking_at_viewer

Jewelry + forehead gem:

jewelry, necklace, pendant, choker, red_gemstone, forehead_mark

*(The forehead gem is weakly trained. If it doesn't render, bump the weight:

(red_gemstone:1.2))*

Outfit 1 — DCAMU leather/club (most reliable, ~70% of training):

`leather_jacket, black_jacket, cropped_jacket, open_jacket, black_dress, high_heel_boots,

black_gloves, fingerless_gloves`

Outfit 2 — purple skirt street look:purple_skirt, pleated_skirt, miniskirt, cropped_jacket, fingerless_gloves, blue_shirt

Setting — bar / nightlife:

bar_(place), bar_stool, bottle, alcohol, night, city_lights, window

Setting — bedroom (NSFW):

on_bed, bed, lying, panties, black_panties, underwear

## Recommended negative additions

patreon_username, web_address, artist_name, watermark, signature, copyright_name

The training set included some images with artist-credit watermarks; without

these in negative they can occasionally leak into outputs.

## Recommended generation settings

- Sampler: DPM++ 2M (my own default) or Euler a

- Schedule: Karras

- CFG: 2.0 (my own default) to 5.5 (Illustrious community default — both work)

- Steps: 20–25

- Resolution: 832×1216 portrait or 1024×1024

- LoRA weight: 0.7–1.0 (1.0 = strongest likeness, 0.7 = lets the base style breathe)

- Clip skip: 2

- Face Detailer / ADetailer: strongly recommended — close-up likeness improves dramatically

## Sample prompt — to reproduce the showcase images

Positive:

r4ven_dc, 1girl, solo, short_hair, bob_cut, cel_shading, flat_color, cartoon, looking_at_viewer,

jewelry, necklace, pendant, red_gemstone, forehead_mark, leather_jacket, black_jacket,

cropped_jacket, open_jacket, black_dress, high_heel_boots, black_gloves, fingerless_gloves,

indoors, window, night, city_lights, masterpiece, best quality, very aesthetic

Negative:

patreon_username, web_address, artist_name, watermark, signature, copyright_name, lowres, worst

quality, bad anatomy, sketch, jpeg artifacts, multiple views, multiple girls, multiple boys

## ⚠️ What this LoRA does NOT cover

This LoRA was trained on 49 images of Raven's DCAMU civilian appearance only —

leather jacket / dress nightlife look and purple-skirt street look. It does

not include her classic Teen Titans hero costume (cloak, hood, leotard).

Prompting cloak, hood, leotard will fall back to base-model interpretation,

not the trained likeness. For the hero look, you'll need a different LoRA.

## Tested on

- Illustrious-XL v0.1 (training base)

- (add any other checkpoints you've tested it on — Nova v18, your Nova3D_0.6 merge, etc.)

## What was trained

49 canon images from DCAMU productions covering two distinct outfits:

- Leather jacket + black dress + high heel boots (urban/nightlife — bulk of dataset)

- Purple pleated skirt + cropped jacket + fingerless gloves (street/casual)

Various poses, expressions, and indoor/outdoor settings (bars, cityscape windows,

parks, amusement park scenes).

## Training details

- Trainer: kohya sd-scripts (LoRA Forge config wrapper)

- Base model: Illustrious-XL v0.1

- Network type: standard LoRA (UNet only)

- Network rank/dim: 32

- Network alpha: 16

- Optimizer: AdaFactor (scale_parameter=False, relative_step=False, warmup_init=False)

- UNet LR: 4e-4

- LR scheduler: cosine_with_restarts (3 cycles)

- LR warmup steps: 100

- Mixed precision: bf16

- Save precision: bf16

- Attention: SDPA (xformers disabled for Blackwell compatibility)

- Resolution: 1024 (bucketed 512–1536, step 64)

- Train batch size: 4

- Dataset: 49 images, 7 repeats per epoch

- Max epochs: 27

- Seed: 42

- Noise: multires_noise_iterations=6, discount=0.3, min_snr_gamma=5,

debiased_estimation_loss=true

- Hardware: RTX 5090