Illustrious & NoobAI: Characters, Prompting, LoRA Training

TL;DR

• Illustrious XL (v2.0 STABLE) and NoobAI XL share a lineage. NoobAI fine-tunes Illustrious on Danbooru + e621. Both recognize thousands of anime characters natively.
• Prompt syntax: write a short English sentence, then append comma-separated Danbooru tags (1girl, red hair, school uniform). NoobAI rewards quality/aesthetic tags (masterpiece, best quality).
• Character LoRA starting point: UNET LR 0.0003, TE LR 0.00003 (10:1 ratio), Dim 64 / Alpha 32, AdamW8bit optimizer. For style LoRAs: UNET LR 0.00025, TE LR 0.00007, Dim 40 / Alpha 20.
• Dataset: 55–60 clean, consistent images minimum. Tag with CivitAI's built-in tagger or JoyTag, then verify every caption.
• Avoid: Prodigy optimizer (unstable on Illustrious/Pony), AI-Toolkit (community prefers kohya_ss), and using an artist's real name as a trigger word for style LoRAs.

What These Models Are

Illustrious XL is an open-source SDXL derivative from OnomaAI Research. Unlike base SDXL (capped at 1024×1024), it generates natively at 1536×1536 and understands a hybrid prompt format: natural English plus Danbooru2023 tags. The project has iterated from v0.1 through v3.5 VPred; v2.0 STABLE introduced cosine annealing for more predictable rendering and is the version most LoRA trainers use as a base. For more on model versions, see internal version history.

NoobAI XL comes from Laxhar Dream Lab. It takes Illustrious and continues training on full Danbooru + e621 tag sets, shipping both ε-prediction (classic SDXL workflow) and v-prediction (richer color, sampler-specific) lines. The community reports broader artist-tag recall and stronger character coverage when you steer with style tags rather than relying on a baked-in look.

Which Characters Work Without a LoRA

The Civitai article "Illustrious XL (v0.1) Recognized Characters List" tested characters across franchises — Marvel, DC, Pokémon, Digimon, Mario, Zelda, Overwatch, Elden Ring, Transformers, Star Wars, Dragon Ball, Star Fox, Undertale, Beastars, League of Legends, and more — by prompting character_name (franchise) on WAI-illustrious-SDXL and Hoseki checkpoints. It color-codes recognition:

Examples reported as Green: Spider-Man, Iron Man, Deadpool, Superman, Batman, Venom, Mario, Link, Ganondorf, Optimus Prime, Charizard, Lucario, Rayquaza, Kratos, The Mandalorian, Boba Fett. Pokémon and Digimon need an extra species tag: pokemon (creature) or digimon (creature).

The author notes that because Illustrious trains on a more recent dataset than Pony Diffusion XL, characters recognized by Pony are likely also recognized by Illustrious — but this is an inference, not a tested equivalence. The Pony list itself (Civitai article 5102) is larger and better maintained; Ponychar.com tracks it programmatically.

Prompt Syntax and Tag Order

Illustrious XL (Hybrid Prompting)

1. Descriptive sentence — one or two clauses: A girl sitting in a cafe, sunlight through the window
2. Danbooru tags — comma-separated specifics: 1girl, red hair, school uniform, coffee, blush
3. Quality anchors (optional but common): masterpiece, best quality, highres

The model interprets both simultaneously. Tag order within the Danbooru block doesn't strictly matter for comprehension, but many users lead with subject (1girl), then appearance, then setting.

NoobAI XL (Tag-First Workflow)

NoobAI documentation explicitly recommends leading with quality/aesthetic tags:

masterpiece, best quality, 1girl, red hair, school uniform, cafe, sunlight

Then add character and artist tags for style control. The ε-pred line behaves like standard SDXL; the v-pred line needs v-prediction-aware settings.

Note: Exact v-pred sampler names should be found in the NoobAI official documentation or tested in practice, as v-prediction requires samplers configured for the v-pred architecture.

Negative Prompts (Both)

Standard anime negative embeddings work:

worst quality, low quality, bad anatomy, watermark, censored, jpeg artifacts, signature, fewer digits, extra digits, bad hands

For character LoRA training, add style conflicts you don't want baked in: 3d, cgi, render, plastic skin, glossy, ((3d model)).

LoRA Training Settings That Work

Settings below are drawn from two community guides (Civitai article 9148 and Civitai article 31875). Both use kohya_ss; one author explicitly advises against AI-Toolkit. Treat these as starting points — the only way to know what works for your dataset is to train, inspect, and adjust. For more training guides, see internal training resources.

Character LoRA (General)

File size: Dim 64 / Alpha 32 ≈ 70 MB (based on Dim 64/Alpha 32 LoRA at SDXL scale per ^[9148]). For complex characters (armor, intricate outfits), some trainers drop to Dim 48 / Alpha 24.

Style / Artist LoRA

File size: Dim 40 / Alpha 20 ≈ 272 MB (reported by ^[31875]; not independently verified).

Trigger word: Do not use the artist's real name — base models already associate it with a style. Invent a unique token (e.g., style_of_artistname) or add descriptors like comic, flat color to captions where they apply.

Optimizer Notes

• AdamW8bit: Memory-efficient, precise LR control. Recommended by both sources. ^[9148], ^[31875]
• AdaFactor (CivitAI on-site default): Cosine with restart, UNET LR 0.0005–0.0003, TE LR 0.00005–0.00003. ^[9148]
• CAME / RAWR: UNET LR 0.0002, TE LR 0.00002, 2500–3000 steps. ^[9148]
• Prodigy: Multiple reports of poor results on Illustrious and Pony; avoid unless you have a specific reason. ^[9148]

Dataset Preparation

1. Count: 55–60 images minimum for style LoRAs; more is better. ^[31875]
2. Sources (ranked by community preference): Twitter/X → Pixiv → Gelbooru → Rule34.us → hitomi.la / imhentai / pawchive for bulk downloads. Avoid rate-limited sites (e621, Rule34.xxx) unless you have tooling.
3. Quality filter: No blur, no crops, no clutter, no style outliers. "Dataset congruence is the biggest factor." ^[31875]
4. Tagging: Use CivitAI's on-site tagger (free) or JoyTag (some prefer it). Export captions as .txt files alongside images.
5. Manual pass: Mandatory. Review every caption. Bad tags bake into the LoRA and are hard to remove.
6. Negative captions (character LoRAs): Add a [Negative Prompt] line to each .txt with conflicting styles (3d, cgi, render...) so they don't leak into the adapter.

Common Errors and Fixes

FAQ

| LoRA file unexpectedly large | Reduce Network Dim; set Alpha to half of Dim. See Dim/Alpha table above. |

| Style LoRA doesn't match the intended artist | Use a unique trigger token; add style descriptors (comic, flat color) to captions. ^[31875] |

| Generations ignore Danbooru tags | Check prompt format: Illustrious uses sentence + tags; NoobAI prefers quality tags first. |

How does NoobAI XL differ from Illustrious XL? NoobAI is a fine-tune of Illustrious trained on Danbooru + e621 tags. It offers ε-pred and v-pred lines, deeper artist-tag recall, and rewards quality tags at prompt start. Illustrious is the more neutral base canvas; NoobAI is the "opinionated" derivative.

What tag order should I use for Illustrious prompting? Sentence first, then Danbooru tags: A girl in a cafe, 1girl, red hair, school uniform, coffee. NoobAI prefers quality tags first: masterpiece, best quality, 1girl, red hair...

How do I train a character LoRA for Illustrious XL? Base: Illustrious v2.0 STABLE. 55–60 clean images. Tag with CivitAI tagger, verify manually. kohya_ss, AdamW8bit, UNET LR 0.0003, TE LR 0.00003, Dim 64 / Alpha 32, batch 1–4, ~1000–3000 steps. Add negative captions for style conflicts.

Can I use Illustrious LoRAs on NoobAI XL? LustrousMix v2 (Hoseki) was merged with NoobAI models for compatibility. Community reports suggest most Illustrious LoRAs transfer, but v-pred NoobAI may need strength adjustment. Test before production. See LoRA compatibility guide for details.

What's the minimum dataset size for a usable LoRA? 55–60 images for style LoRAs per ^[31875]. Character LoRAs: no consensus minimum; use 55+ images for consistency. More data generally improves fidelity, but rank and training steps matter more than raw count alone.

Why does my LoRA look like 3D / plastic? 3D-style images slipped into the dataset or negative captions weren't used. Add 3d, cgi, render, plastic skin, glossy, ((3d model)) to per-image negative captions during training.

Sources

• Illustrious XL official site — model architecture, versions, prompting system, licensing
• NoobAI XL official site — lineage, tag coverage, ε-pred/v-pred lines, usage guide
• Illustrious XL v2.0 on Hugging Face — cosine annealing checkpoint details
• WAI-illustrious-SDXL on Civitai — checkpoint used for character testing
• Hoseki on Civitai — LustrousMix/Anima checkpoint used for character testing
• Illustrious XL character list (Civitai article 10242) — community recognition tiers, prompt method
• Illustrious LoRA training discussion (Civitai article 9148) — LR ranges, Dim/Alpha, optimizer comparisons
• Yeetyah's training guide (Civitai article 31875) — dataset prep, style LoRA sweet spots, optimizer advice
• Pony Diffusion XL character list (Civitai article 5102) — referenced for cross-model comparison