stable-diffusion-3.5-medium-HD

This is a PEFT LoRA derived from stabilityai/stable-diffusion-3.5-medium.

The main validation prompt used during training was:

Designed by Hyundai, front view, compact sedan, three-box silhouette, long wheelbase, wide track, short overhangs, low stance, taut body surfacing, aggressive front fascia, mesh grille, large air intakes, swept-back DRL signature, LED headlamps, flush door handles, multi-spoke 18-inch wheels, low-profile tires, chrome window trim, low beltline, roof rails, metallic gray paint, gloss black accents

Validation settings

• CFG: 7.5
• CFG Rescale: 0.0
• Steps: 20
• Sampler: FlowMatchEulerDiscreteScheduler
• Seed: 42
• Resolution: 1024x1024
• Skip-layer guidance:

Note: The validation settings are not necessarily the same as the training settings.

You can find some example images in the following gallery:

Prompt

unconditional (blank prompt)

Negative Prompt

blurry, cropped, ugly

Prompt

Designed by Hyundai, front view, compact sedan, three-box silhouette, long wheelbase, wide track, short overhangs, low stance, taut body surfacing, aggressive front fascia, mesh grille, large air intakes, swept-back DRL signature, LED headlamps, flush door handles, multi-spoke 18-inch wheels, low-profile tires, chrome window trim, low beltline, roof rails, metallic gray paint, gloss black accents

Negative Prompt

blurry, cropped, ugly

The text encoder was not trained. You may reuse the base model text encoder for inference.

Training settings

• Training epochs: 2

• Training steps: 1074

• Learning rate: 0.0001

  • Learning rate schedule: sine
  • Warmup steps: 500

• Max grad value: 2.0

• Effective batch size: 1

  • Micro-batch size: 1
  • Gradient accumulation steps: 1
  • Number of GPUs: 1

• Gradient checkpointing: True

• Prediction type: flow_matching (extra parameters=['shift=3'])

• Optimizer: adamw_bf16

• Trainable parameter precision: Pure BF16

• Base model precision: no_change

• Caption dropout probability: 0.0%

• LoRA Rank: 256

• LoRA Alpha: 256.0

• LoRA Dropout: 0.1

• LoRA initialisation style: default

• LoRA mode: Standard

Datasets

test_dataset

• Repeats: 1
• Total number of images: 179
• Total number of aspect buckets: 1
• Resolution: 1.048576 megapixels
• Cropped: False
• Crop style: None
• Crop aspect: None
• Used for regularisation data: No

Inference

import torch
from diffusers import DiffusionPipeline

model_id = 'stabilityai/stable-diffusion-3.5-medium'
adapter_id = 'mingyu-oo/stable-diffusion-3.5-medium-HD'
pipeline = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) # loading directly in bf16
pipeline.load_lora_weights(adapter_id)

prompt = "Designed by Hyundai, front view, compact sedan, three-box silhouette, long wheelbase, wide track, short overhangs, low stance, taut body surfacing, aggressive front fascia, mesh grille, large air intakes, swept-back DRL signature, LED headlamps, flush door handles, multi-spoke 18-inch wheels, low-profile tires, chrome window trim, low beltline, roof rails, metallic gray paint, gloss black accents"
negative_prompt = 'blurry, cropped, ugly'

## Optional: quantise the model to save on vram.
## Note: The model was not quantised during training, so it is not necessary to quantise it during inference time.
#from optimum.quanto import quantize, freeze, qint8
#quantize(pipeline.transformer, weights=qint8)
#freeze(pipeline.transformer)
    
pipeline.to('cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu') # the pipeline is already in its target precision level
model_output = pipeline(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=20,
    generator=torch.Generator(device='cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu').manual_seed(42),
    width=1024,
    height=1024,
    guidance_scale=7.5,
).images[0]

model_output.save("output.png", format="PNG")