SDXL Checkpoints — Juggernaut XL v9 + Illustrious Anime v4

Full single-file SDXL checkpoints, ready to load as-is.

What this repo is: two complete single-file SDXL checkpoints plus the shared tokenizers — weights only, not a retrain. The recommended settings below are the values these weights are actually run with day to day, not copied from the upstream card.

Samples

All samples below add this negative prompt, which is not part of the preset in the settings table:

washed out, low contrast, hazy, faded, pastel, bloom, overexposed, glare, milky, foggy, flat lighting, blown highlights, colour cast

Everything else — resolution, steps, guidance, seed — is exactly the documented preset. Without it these checkpoints, Illustrious especially, render with a heavy pastel bloom; see the note on it below.

img2img

Both checkpoints do img2img as well as txt2img. Two sources, each through both checkpoints.

Describe the image you want. Do not give an instruction.

SDXL has no instruction-following training. The prompt is a description of the image to produce, not a command applied to the source:

✅ a still life oil painting of a red apple on a rustic wooden table beside a sunlit window…

❌ the same scene rendered as a vivid oil painting…

The instruction form is taken literally. Asking for "the same scene rendered as an oil painting" returned an apple coated in dripping paint with a paintbrush in frame, and at higher strength a different scene altogether — at the same strength 0.75 that produced the top row here.

The two checkpoints

Both keep the subject at 0.75 — same apple, position, scale and light. Juggernaut is heavier impasto and warmer; Illustrious is cleaner and brighter (retaining ~93% of source luminance against Juggernaut's ~85%), rendering more as illustration.

strength and the swept range

Swept 15–40 steps and guidance 5.0–12.5 at strength 0.75, output was stable throughout. Below ~0.3 gives little more than a gloss pass; above ~0.85 the source becomes a loose suggestion.

On the synthetic source

Bottom row: Juggernaut 1.00× source luminance, Illustrious 1.03×. Juggernaut applies vertical brushwork to the gradient; Illustrious leaves it close to the source. In both, the flat circle is reproduced flat — a uniform region carries no detail for the sampler at strength 0.75, while the graduated area does. Two consequences for testing: subject loss cannot be observed against a subjectless source (strength 0.90 passes here and does not preserve a photograph's subject), and histogram_correlation measures 0.001/−0.017 here against 0.73–0.78 for the same runs on a photograph.

Recommended settings

Values this build is run with in practice. They are read out of a live config when this card is generated, so they cannot drift from what is actually used.

Supported modes: txt2img, img2img

Notes and gotchas

• This repo holds two checkpoints and the shared SDXL tokenizers — there is no unet/VAE folder, because each .safetensors here is a complete single-file checkpoint.
• Two independent checkpoints. Each .safetensors here is complete on its own; load whichever you want. They share the tokenizers under config/.
• These are run at float32, so expect a larger memory footprint than an fp16 SDXL deployment.
• Unlike every other model here, SDXL uses a real CFG schedule — guidance 7.5. It is not a distilled/turbo build.
• Expect a pastel bloom unless you negate it. Illustrious in particular renders with a warm pink-lavender haze and lifted blacks — a trained style trait, not a fault: measured against the prompt it is the most saturated output here, it is just low-contrast and bright. A negative prompt naming the artifact (washed out, hazy, pastel, bloom, …) removes it cleanly — on the anime street it moved tonal spread 164 → 208 and dropped mean brightness 161 → 142. Raising guidance also increases contrast but amplifies everything, including the magenta cast, so the negative prompt is the better lever. On the photographic Juggernaut renders the same negative mostly raises colour saturation (apple 70 → 118) rather than removing haze, so judge it per checkpoint rather than applying it blindly.
• OpenCLIP-G needs standard GELU, not QuickGELU — this one silently destroys Illustrious-family checkpoints. SDXL has two text encoders: CLIP-L (OpenAI, QuickGELU) and OpenCLIP-G / bigG (LAION, standard GELU). Defaulting both to QuickGELU corrupts the bigG embeddings while the weights still load cleanly, so nothing in the logs looks wrong. juggernaut_XL_v9 tolerated it and rendered normally; illustrious_Anime_v4 did not, producing pure noise and near-blank canvases (pixel std as low as 4.1) on every attempt. Fixed 2026-07-18; both checkpoints now render correctly at the settings above, and every sample on this card is a post-fix render.

Layout

Each .safetensors under combined/ is a complete SDXL checkpoint — unet, both text encoders and the VAE in one file — so any standard SDXL loader can take it directly. The shared CLIP-L and OpenCLIP-G tokenizers live under config/tokenizer/ and config/tokenizer_2/. There is no per-component split here.

If you load these outside the sidecar, set the OpenCLIP-G text encoder's activation to standard GELU. QuickGELU is the right default for CLIP-L only, and getting it wrong yields noise rather than an error — see the notes above.

Files

Provenance

• Upstream base model: stabilityai/stable-diffusion-xl-base-1.0 (upstream license: openrail++)
• This build: community fine-tunes of SDXL, redistributed as-is with the shared tokenizers alongside. Weights are not retrained or requantized here.
• License: left as unknown in this repo's metadata. Refer to the upstream model's license for redistribution and commercial-use terms — several of these bases are non-commercial.