Gatchamn/juggernaut-x-hyper-fp16

Juggernaut X-Hyper — fp16 diffusers tree

A format conversion of RunDiffusion/Juggernaut-X-Hyper. The weights are RunDiffusion's; nothing has been retrained, merged, or fine-tuned.

Why this repo exists

The upstream repo cannot be loaded by diffusers/MLX as published:

• its diffusers tree stores weights as PyTorch pickle (.bin), not safetensors;
• its tokenizers ship without the legacy CLIP pair (vocab.json / merges.txt) that SDXL text-encoding pipelines expect.

The only directly usable weights upstream are in the root single-file checkpoint, JuggernautXRundiffusion_Hyper.safetensors.

Changes made

1. Loaded the upstream single-file checkpoint with StableDiffusionXLPipeline.from_single_file(...) and re-saved it as a standard fp16 diffusers tree (safe_serialization=True, variant="fp16").
2. Restored vocab.json and merges.txt for tokenizer/ and tokenizer_2/ from stabilityai/stable-diffusion-xl-base-1.0. SDXL fine-tunes do not retrain the CLIP tokenizers, so these files are byte-identical across SDXL models (vocab.json 1,059,962 bytes, merges.txt 524,619 bytes).

No weight values were altered. The conversion script is sdxl_single_file_to_diffusers.py.

Recommended settings

Per the upstream model card:

Setting	Value
Steps	4–8 (start at 6)
CFG scale	1.0–2.0
Sampler	DPM++ SDE or TCD
Resolution	1024×1024 (SDXL native)

Verified on an M-series Mac Studio via MLX: 6 steps, CFG 2.0, 1024×1024, ~7.5s per image.

License

CreativeML Open RAIL-M, inherited unchanged from the upstream model — see the original model card and the license text.

Upstream states: "This model may not be deployed behind paid API services without explicit licensing." Commercial licensing: juggernaut@rundiffusion.com.

The RAIL-M use-based restrictions apply to this copy exactly as they apply to the original.

Attribution

Juggernaut X-Hyper by RunDiffusion / KandooAI. Original: https://huggingface.co/RunDiffusion/Juggernaut-X-Hyper