Stable Diffusion 3.5 Medium vs Large: Setup, Usage and Practical Guide

Stable Diffusion 3.5 is Stability AI's text-to-image model family. This guide focuses on Stable Diffusion 3.5 Medium and Stable Diffusion 3.5 Large, how they differ, what they are useful for, and how to run them with Hugging Face Diffusers or ComfyUI.

Medium is usually the practical choice for faster local inference and lower memory use. Large is better when image quality, difficult compositions, and prompt adherence matter most.

Architecture

SD3.5 builds on the Multimodal Diffusion Transformer (MMDiT) design. The family uses multiple text encoders, including CLIP and T5, to represent prompts with richer semantic information.

Key goals include better complex-prompt understanding, image coherence, typography, flexible resolutions, local deployment, and customization.

Medium vs Large

Medium is well suited to article covers, social graphics, prototypes, game/UI assets, LoRA experiments, and local services. Large is better for professional key visuals, advertising, complicated scenes, detailed concept art, and ControlNet-guided workflows.

Installation

python -m venv .venv

Linux/macOS:

source .venv/bin/activate

Windows:

.venv\Scripts\Activate.ps1

Install:

pip install -U torch diffusers transformers accelerate safetensors sentencepiece huggingface_hub

The official Hugging Face repositories are gated. Accept the model terms and authenticate:

hf auth login

Model IDs:

stabilityai/stable-diffusion-3.5-medium
stabilityai/stable-diffusion-3.5-large

Medium with Diffusers

import torch
from diffusers import StableDiffusion3Pipeline

pipe = StableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3.5-medium",
    torch_dtype=torch.bfloat16,
).to("cuda")

image = pipe(
    "A futuristic AI research lab, blue lighting, cinematic, highly detailed",
    num_inference_steps=28,
    guidance_scale=5.0,
    height=1024,
    width=1024,
).images[0]

image.save("sd35-medium.png")

Large

import torch
from diffusers import StableDiffusion3Pipeline

pipe = StableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3.5-large",
    torch_dtype=torch.float16,
)

pipe.enable_model_cpu_offload()

image = pipe(
    "A brutalist museum floating above a calm ocean at sunset",
    negative_prompt="low quality, blurry, distorted",
    num_inference_steps=28,
    guidance_scale=5.0,
    height=1024,
    width=1024,
).images[0]

image.save("sd35-large.png")

Reducing memory use

Common options are FP16/BF16, CPU offloading, lower resolution, quantization, FP8 workflows where supported, and using Medium instead of Large.

ComfyUI

Place the checkpoint in:

ComfyUI/models/checkpoints/

and text encoders in:

ComfyUI/models/text_encoders/

Typical SD3.5 workflows use:

clip_l
clip_g
t5xxl

An FP8 T5-XXL can reduce memory pressure.

Prompting

Describe subject, environment, lighting, camera, style, and details explicitly. A practical starting range is 25–35 inference steps and a guidance scale around 4–7. Use a fixed seed when reproducibility matters.

ControlNet

Official Canny, Depth, and Blur ControlNets are available for SD 3.5 Large. They provide stronger structural control than prompting alone.

Applications

• blog and website imagery
• advertising
• social graphics
• concept art
• game assets
• product visualization
• LoRA/fine-tuning
• internal image generation services

Production notes

Track inference latency, VRAM/RAM, queueing, batching, caching, GPU cost, safety, model versioning, and reproducibility. Save model ID, prompt, negative prompt, seed, steps, guidance, and resolution as metadata.

License

Both models use the Stability AI Community License. Review the latest license and commercial terms before shipping a commercial product.

Official resources

• Stable Diffusion 3.5 Medium
• Stable Diffusion 3.5 Large
• Introducing Stable Diffusion 3.5
• Diffusers SD3 Pipeline
• SD3.5 Reference Implementation
• ComfyUI SD3.5 Examples