V3LOURS

What this is

For a year I have been making images on Illustrious with a stack of five style LoRAs and a long, fixed style prompt. It works, but it is heavy: five resources to load, a prompt prefix I paste every time, and a look that only exists as long as I keep that whole rig together.

This LoRA is that rig, folded into one file for Anima.

It is not a copy. The point was never to reproduce the same pictures — it was to carry the hand over to a base model that understands sentences, composes scenes, and handles more than one person in frame.

Trigger word: dwil

---

Where the training data comes from

337 of my own published images, from 3 March to 25 August 2026. Nothing scraped, nothing borrowed — every image is one I made and posted. 256 of them are 1248×1824, so the model learned from hi-res-fixed output, not from small drafts. That is why it holds up under a second pass.

The set is honest about what it is: heavily portrait, mostly single figures, mostly night and interior, and it includes an adult range — the source posts run from Soft through X. A small block of 32 images carries no people at all, deliberately, so that the style would not be welded to "a woman looking at the camera".

---

The Illustrious stack it comes from

Credit where it is due — this look was built on other people's work:

0.8 Ch4mpi | Shiiro's Styles

0.7 Modern anime render

0.5 Cartoon styles — Midjourney S2

0.4 Realistic Skin Texture style

0.2 Sinozick | Shiiro's Styles | Niji

With the main checkpoints been IllustriousNXT_XL v2.0,
Most of the images at CFG 4, 29 steps, Euler a.

And this style prefix, in front of every prompt:

Sinozick_illu, gradient, spot color, Carto4on, ch4mpi_illu, modern_anime_render,

high_detail, glossy_highlights, vivid_colors, dynamic_lighting, atmospheric_depth,

soft_shading, painterly_texture, clean_lineart, (flat color:2), (no lineart:1),

(no outline:1), (Flat vector:1), cinematic photography style

dwil replaces all of that. One token, one file, on a different base model.

---

How it was trained

• kohya-ss/sd-scripts, anima_train_network.py, networks.lora_anima

• Base: anima-base-v1.0

• Rank 16, alpha 16 — rank 32 was tried and made no difference I could see

• 4250 steps, 12 epochs, 337 images / 363 views after repeats

• AdamW8bit, LR 1e-4, cosine_with_restarts ×3, bf16, bucketed at 1024

• Captions are booru-style tags, trigger kept in first position

• ~4h50 on a single 8 GB card

---

Recommended settings

• Base: anima-base-v1.0 — weight 1.0

• CFG 4.0, 30 steps, sampler Euler a, schedule Normal

• 832×1216 portrait or 1216×832 landscape

• Hi-res fix ×1.5, denoise 0.5 — this is where it looks best, and it is how the training images were made

Weight behaves smoothly from 0.6 to 1.2. Lower weight keeps more of the base model's contrast and grain; higher weight flattens further. 1.0 is where I work.

---

What it does not do

It flattens. That is the point, but it means this is not a texture or micro-detail LoRA — at full weight it carries less surface grain than the base model, not more. If you want punchy high-contrast output, run it at 0.6 or lighten your prompt.

Wide shots flatten more than close-ups. The training set is portrait-heavy and it shows. Group scenes and complex multi-character compositions are not what it was built for, and I have not verified that it helps there.

One quirk worth knowing, since it is baked into the source style: my Illustrious prompt asks for clean_lineart and (no lineart:1) at the same time. Those two fight each other, and the flatness you see is partly the outcome of that fight.

---

Measured, not claimed

Every number above comes from paired measurements — same prompts, same seeds, one factor changed at a time. The gallery includes wide shots, low light, a non-human subject and a six-character scene precisely because those are the hard cases, not because they all succeeded.