### About the Model
v2 is a complete retrain. Where v1 was a "CGI enhancer" that worked best stacked on top of other 3D LoRAs, v2 is the renderer itself: use it alone, at full weight, and it takes over how the image is lit and shaded.
The target look is a modern offline raytraced character render — Cycles / Octane / RE Engine class: real sub-surface scattering on skin, visible veins, sweat and wet specular highlights, physically based lighting with deep shadows and practical light sources, sharp fabric and hair micro-detail. Anatomy and character features stay yours; the LoRA only decides how they are rendered.
Trained on 1,314 hand-curated images, all from the same render language — no cel-shaded / game-CG / Koikatsu / MMD mixed in, which is exactly what kept v1 halfway between anime and 3D. Sources were selected by artist rather than by tag, re-fetched at full resolution (1400 px+), signature-cropped and de-duplicated. It is an NSFW-first dataset (~94% explicit/nsfw); SFW works, but that's not where the data is.
Works with the Anima turbo LoRA.
---
### How to Use & Prompting Strategy
* Primary Trigger Word: @lllsuperlllcgi (same as v1)
* Style Anchors (Required): 3d, realistic, photo (medium) — these were in every training caption and carry part of the style. Put them right after the trigger.
* Rating tag second: start the prompt with newest, explicit / nsfw / sensitive / safe. The set is mostly explicit and nsfw; those two also push the render hardest.
* Describe, don't name. Characters were captioned by appearance only (hair, eyes, outfit), never by name — so describe the look; character LoRAs stack fine.
* Do not stack v1 or other style LoRAs. Do not use (blender \(medium\)) or cgi anchors from v1 — they were deliberately left out of v2 captions.
Prompt skeleton:
> newest, explicit, @lllsuperlllcgi, 3d, realistic, photo (medium), 1girl, brown hair, ponytail, green eyes, freckles, nude, sitting on bed, dim bedroom, warm lamp light, sweat, looking at viewer
---
### Recommended Generation Settings
* LoRA Weight: 1.0 (0.8–1.0). This is not an enhancer — run it at full strength.
* Base Model: WAI-ANIMA v1.0 (trained on it). Anima v1.0 Base also works.
* Sampler: er_sde + beta for maximum skin/fabric micro-detail; euler a + normal for a cleaner, softer CGI look — the style holds on both.
* Steps: 28 – 35
* CFG Scale: 3.0 – 3.5 (keep it low — flow model)
* Shift: 3.0
* Resolution: 832×1216 / 1024×1024 and up
* Negative Prompt:
> worst quality, low quality, blurry, jpeg artifacts, lowres, watermark, signature, text, censored, bad hands, anime, flat color
Amateur Filter from v1 (sharpen 0.01 / desaturation 0.85 / grain 0.02) still helps on euler a to kill the last of the AI sheen; on er_sde it is usually unnecessary.
---
### Training (v2)
* Base: WAI-ANIMA v1.0 · LoRA rank 32 / alpha 32 · LR 1e-4, bf16
* 1,314 images at 1024 with aspect buckets · batch 4 · 8 epochs = 2,712 steps
* Captions: WD ViT-Large tagger v3 + trigger, 6 pinned tokens, shuffled, 10% tag dropout
* Text encoder and LLM adapter frozen
* Three configs were trained side by side (LoKr r64 @1e-4, LoKr r64 @2e-4, LoRA r32 @1e-4) and compared on a fixed grid against v1 on both samplers; LoKr @2e-4 diverged mid-run. This release is the LoRA r32 run at epoch 8 — the most stable trajectory, lowest final loss, best prompt adherence.
---
### v1 vs v2
* Role — v1: enhancer, stack with other 3D LoRAs → v2: standalone renderer
* Weight — v1: 0.7–0.9 → v2: 1.0
* Dataset — v1: ~120 images, preview resolution, mixed styles → v2: 1,314 images, 1400 px+, one render language
* Anchors — v1: cgi, (blender \(medium\):0.7) → v2: 3d, realistic, photo (medium)
* Best sampler — v1: DPM2 / Euler + Karras → v2: er_sde + beta / euler a + normal
Description
## About v2
v1 was a flexible "CGI enhancer": great stacked on top of other 3D LoRAs, but on its own it landed halfway between anime and a Blender render. v2 is a full retrain from a new dataset and is meant to be used on its own, at weight 1.0, as the renderer — not as a booster. It also works with the Anima turbo LoRA.
What changed:
* New dataset, 1,314 images, all one render language: offline raytraced (Cycles/Octane-class) character renders with real sub-surface scattering on skin, veins, sweat and wet specular, physically based lighting — dark interiors, single-source practical lights, rim light. Nothing cel-shaded, no Koikatsu / Honey Select / MMD / VRoid, no game screenshots, no upscales.
* Sourced by artist, not by tag. The 3d tag on boorus is mostly SFM/Koikatsu. Instead the set was built from a handful of 3D artists whose renders share the same look, plus a hand-picked set from v1 re-fetched at full resolution (v1 had been trained on ~1080 px previews — one of the reasons it stayed soft).
* Minimum 1400 px short side on source, cropped 5% per side to remove signatures, trained at 1024 with aspect buckets. Near-duplicates removed by perceptual hash. Max 60 images per named character so the LoRA learns rendering, not faces.
* Captions describe, never name. No character names, no artist names, no blender (medium). Trigger + medium tags pinned first, WD-tagger tags after, shuffled with 10% dropout. Explicit content is tagged normally so prompt control stays intact.
* Rating split of the set: explicit ~77%, nsfw ~17%, sensitive ~6%. It is an NSFW-first LoRA; SFW works, but that's not where the data is.
## Trigger & prompting
```
newest, explicit, @lllsuperlllcgi, 3d, realistic, photo (medium), <your scene>
```
* @lllsuperlllcgi — trigger (same as v1).
* 3d, realistic, photo (medium) — were in every caption; the style partly lives on them, keep them.
* Second token is the rating: explicit / nsfw / sensitive / safe — the set is mostly explicitnsfw, and those two also push the render harder.
* Describe characters by appearance (hair, eyes, outfit). Names were not trained.
* Don't stack v1 or other style LoRAs. Character LoRAs and the turbo LoRA are fine.
Weight: 1.0 (0.8–1.0 range).
## Recommended settings
Base: *WAI-ANIMA v1.0** (trained on it) — Anima base also works.
* Sampler: er_sde + beta for the sharpest skin/fabric micro-detail; euler a + normal gives a softer, cleaner CGI look and holds the style too.
* CFG 3–3.5, shift 3, 28–35 steps, 832×1216 / 1024×1024.
* Negative: worst quality, low quality, blurry, jpeg artifacts, lowres, watermark, signature, text, censored, bad hands, anime, flat color
## Training
| | |
|---|---|
| Base | WAI-ANIMA v1.0 (base 1.0) |
| Type | LoRA, rank 32 / alpha 32 |
| Resolution | 1024, aspect buckets (29 buckets) |
| Batch | 4 |
| LR | 1e-4, bf16 |
| Steps | 2,712 (8 epochs × 339 steps) |
| Dataset | 1,314 images, 1400 px+ sources |
| Text encoder / LLM adapter | frozen |
Three configs were trained side by side (LoKr r64 at LR 1e-4, LoKr r64 at LR 2e-4, LoRA r32 at LR 1e-4) and compared on a fixed 6-prompt grid, same seed, on both euler a and er_sde, against v1. LoKr at 2e-4 diverged mid-run (black frames at epoch 4, recovered by 8). The released file is the LoRA r32 run at epoch 8: the most stable trajectory of the three, lowest final loss, and the best prompt adherence on the grid.
## Notes if you train your own Anima style LoRA
* Source by artist, cap by character, describe instead of naming.
* Don't train the LLM adapter; keep LR at 1e-4 or lower — 2e-4 on LoKr blew up.
* Train at 1024 from ≥1400 px sources. Previews make soft LoRAs.
* Crop signatures; a corner logo is in every generation by epoch 3.
* With Anima's tag captioning, batch >1 needs bucket-aware batching (the stock toolkit stacks mismatched latents).