illustrious (JukeBOX)

An anime and illustration checkpoint for character portraits, fantasy adventures, cinematic city scenes and colorful everyday moments. Jukebox brings WAI Illustrious to SDXL with his custom training and good eye to anime detail!

Think expressive faces, illustrated hair and clothing, clean anime-style shading, soft light and richly described settings. Try a quiet café portrait, an armored adventurer in an ancient hall, or a rain-soaked city full of neon. These are creative directions to explore, not guarantees: the prompt, seed, resolution and chosen format all influence the result.

## What this release is

It retains the SDXL/Illustrious architecture, both text encoders, and an integrated VAE. No trigger word is required by the merge itself.

## Suggested starting settings

• Steps: 25–32

• CFG scale: 4–7.5; start around 5.5–6

• Sampler: Euler a, or Euler with the Simple scheduler

• Resolution: start at 1024 × 1024, 832 × 1216 or 1216 × 832

• Batch size: 1 while establishing your settings

• VAE: included; use the checkpoint's VAE

• CLIP: use your application's SDXL/Illustrious default penultimate-layer setting; this is commonly described as Clip skip 2, but UI labels differ

If your interface separates the sampler from its scheduler, “Euler / Simple” means sampler = Euler and scheduler = Simple. Euler a is a separate ancestral sampler.

Hires fix starting point:

• Upscale: about 1.5× to begin

• Upscaler: R-ESRGAN 4x+ Anime6B

• Hires steps: around 15–20

• Denoising strength: 0.35–0.50

The “4x” in the upscaler's name does not require a 4× final image. Set the Hires scale separately. Lower denoising generally keeps the original composition closer; higher denoising allows more reinterpretation. Reduce the scale or disable Hires fix if memory is tight. These are starting settings, not measured optimal settings for this merge.

## Prompting with Danbooru / Gelbooru-style tags

Use a focused, comma-separated description. A useful order is: subject and count → appearance → clothing → pose or action → framing → setting → lighting → style. Add a short natural-language sentence when you need to describe a relationship between subjects or a specific composition.

For example:

1girl, adult woman, solo, short black hair, brown eyes, red jacket, upper body, looking at viewer, city street, rain, neon lights, rim lighting, anime coloring

Danbooru and Gelbooru use searchable tag vocabularies. They are useful for finding concise descriptions, but a tag existing on either site does not guarantee that this checkpoint understands it. Their vocabularies overlap rather than matching perfectly. This release does not claim separate Gelbooru training.

• Translate descriptive tags into ordinary prompt phrases where appropriate: long_hair → long hair; looking_at_viewer → looking at viewer.

• Keep recognized character names intact if you use them. Parentheses inside a name may need escaping in WebUI/Forge, because unescaped parentheses are also used for prompt emphasis.

• Do not paste website search operators such as score:>50, order:score or uploader:name into the image prompt; they filter website searches, not generated images.

• Choose one framing direction: portrait, upper body or full body. Combining contradictory framing requests makes the prompt less clear.

• Start with a short prompt. Add a few details at a time while holding the seed fixed.

• If your interface supports it, a modest weight such as (red jacket:1.1) can emphasize an important detail. That weighting syntax belongs to the generation interface, not to Danbooru/Gelbooru.

• Avoid long lists of quality synonyms and enormous negative prompts. WAI's upstream guidance specifically discourages that. Start with a few useful quality terms and remove ones that do not help.

• For non-explicit images, state that clearly in the prompt, use appropriate negative tags and inspect the result. Negative tags alone are not a content guarantee.

## Credits, recipe and terms

WAI Illustrious over many years of use we developed our own brand and type of illustrious model, keeping the same flow feel and core of the original, as best as we could.

Illustrious foundation: Onoma AI Research

https://huggingface.co/OnomaAIResearch/Illustrious-xl-early-release-v0

CLIP-L repair reference: Stability AI SDXL Base 1.0

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

The upstream Illustrious card specifies Fair AI Public License 1.0-SD.