metadata

language:
  - en
library_name: diffusers
inference: true
license: other
license_name: stabilityai-ai-community
license_link: LICENSE.md
tags:
  - text-to-image
  - stable-diffusion
  - diffusers
base_model:
  - stabilityai/stable-diffusion-3.5-large
  - stabilityai/stable-diffusion-3.5-large-turbo
base_model_relation: merge

Stable Diffusion 3.5 Merged

This repository contains the merged version of Stable Diffusion 3.5, combining the best features from both the Large and Turbo variants.

Large (40 steps)	Turbo (4 steps)	Merged (6 steps) 🎉

Inference

Run the following code to generate images using the merged model:

from diffusers import StableDiffusion3Pipeline
import torch

pipeline = StableDiffusion3Pipeline.from_pretrained(
    "ariG23498/sd-3.5-merged", torch_dtype=torch.bfloat16
).to("cuda")

prompt = "a tiny astronaut hatching from an egg on the moon"
image = pipeline(
    prompt=prompt,
    guidance_scale=1.0,
    num_inference_steps=6,  # Run faster ⚡️
    generator=torch.manual_seed(0),
).images[0]
image.save("sd-3.5-merged.png")

Note: Turbo variant runs faster with fewer steps, while Large variant requires more steps (around 50) but provides better detail. With the merged model you would need to play with num_inference_steps and guidance_scale to get the perfect balance of speed and quality. Below I show a grid of scale and step changes and its corresponding generations.

Merging Models

This repository merges the Stable Diffusion 3.5 Large and Stable Diffusion 3.5 Turbo models into a single, powerful model. The Large version uses classifier-free guidance (CFG) and requires more steps, while the Turbo version is distilled for faster generation without CFG.

The merged model retains the detail of the Large version and the speed of the Turbo version.

Code to Merge Models

To access the Stable Diffusion 3.5 models, one needs to fill the forms in the corresponding repositories, and then huggingface_cli login to let your system know who you are and whether you have access to the models!

from diffusers import SD3Transformer2DModel
from huggingface_hub import snapshot_download
from accelerate import init_empty_weights
from diffusers.models.model_loading_utils import load_model_dict_into_meta
import safetensors.torch
from huggingface_hub import upload_folder
import glob
import torch

large_model_id = "stabilityai/stable-diffusion-3.5-large"
turbo_model_id = "stabilityai/stable-diffusion-3.5-large-turbo"

with init_empty_weights():
    config = SD3Transformer2DModel.load_config(large_model_id, subfolder="transformer")
    model = SD3Transformer2DModel.from_config(config)

large_ckpt = snapshot_download(repo_id=large_model_id, allow_patterns="transformer/*")
turbo_ckpt = snapshot_download(repo_id=turbo_model_id, allow_patterns="transformer/*")

large_shards = sorted(glob.glob(f"{large_ckpt}/transformer/*.safetensors"))
turbo_shards = sorted(glob.glob(f"{turbo_ckpt}/transformer/*.safetensors"))

merged_state_dict = {}
guidance_state_dict = {}

for i in range(len((large_shards))):
    state_dict_large_temp = safetensors.torch.load_file(large_shards[i])
    state_dict_turbo_temp = safetensors.torch.load_file(turbo_shards[i])

    keys = list(state_dict_large_temp.keys())
    for k in keys:
        if "guidance" not in k:
            merged_state_dict[k] = (state_dict_large_temp.pop(k) + state_dict_turbo_temp.pop(k)) / 2
        else:
            guidance_state_dict[k] = state_dict_large_temp.pop(k)

    if len(state_dict_large_temp) > 0:
        raise ValueError(f"There should not be any residue but got: {list(state_dict_large_temp.keys())}.")
    if len(state_dict_turbo_temp) > 0:
        raise ValueError(f"There should not be any residue but got: {list(state_dict_turbo_temp.keys())}.")

merged_state_dict.update(guidance_state_dict)
load_model_dict_into_meta(model, merged_state_dict)

model.to(torch.bfloat16).save_pretrained("transformer")

upload_folder(
    repo_id="ariG23498/sd-3.5-merged",
    folder_path="transformer",
    path_in_repo="transformer",
)

This script downloads the checkpoints, merges them, and saves the merged model locally. You can then upload the merged model to Hugging Face Hub using upload_folder.

References:

FLUX.1 merged from Sayak Paul!