phobos36/sdxl-base-1.0-single

SDXL Base 1.0 — minimal Diffusers package

This main branch is a complete Diffusers multifolder layout for Model RUN and StableDiffusionXLPipeline.from_pretrained(). It mirrors the official stabilityai/stable-diffusion-xl-base-1.0 components at revision 462165984030d82259a11f4367a4eed129e94a7b. It is not an official Stability AI repository.

The earlier single-file ComfyUI checkpoint is preserved on the comfy-checkpoint branch. Main intentionally has no combined checkpoint, ONNX, Flax, duplicate precision variants, or refiner. The model is split into the components required by Diffusers.

Layout and precision

model_index.json
scheduler/scheduler_config.json
unet/config.json
unet/diffusion_pytorch_model.safetensors       # official FP16 weights
text_encoder/config.json
text_encoder/model.safetensors               # official CLIP-L FP16 weights
text_encoder_2/config.json
text_encoder_2/model.safetensors              # official CLIP-G FP16 weights
tokenizer/                                  # vocab, merges, tokenizer configs
tokenizer_2/                                # vocab, merges, tokenizer configs
vae/config.json
vae/diffusion_pytorch_model.safetensors      # official FP32 weights

All four weight files are copied byte-for-byte from upstream; hashes are listed in SHA256SUMS. FP16 source files use the standard, unsuffixed Diffusers filenames so loading does not require a variant="fp16" argument. VAE weights stay FP32 and their original force_upcast setting is retained.

import torch
from diffusers import StableDiffusionXLPipeline

pipe = StableDiffusionXLPipeline.from_pretrained(
    "phobos36/sdxl-base-1.0-single", torch_dtype=torch.float16, use_safetensors=True,
)
pipe.load_lora_weights(
    "phobos36/sdxl-lightning-4step-lora",
    weight_name="sdxl_lightning_4step_lora.safetensors",
)
pipe.fuse_lora()
pipe.to("cuda")
image = pipe("A girl smiling", num_inference_steps=4, guidance_scale=0).images[0]

Lightning is not merged into the base weights. The adapter is a separate repository. The scheduler configuration uses EulerDiscreteScheduler with timestep_spacing="trailing" and prediction_type="epsilon" as required by SDXL-Lightning. All other upstream component configuration files are preserved unchanged.

Cloud.ru Model RUN should use runtime Diffusers and task text-to-image. This package has not yet been run on Cloud.ru or tested on a GPU.

Copyright (c) 2023 Stability AI. The original CreativeML Open RAIL++-M license, including its use restrictions, is preserved in LICENSE.md.