sd35m-reflow

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

The main validation prompt used during training was:

A photo-realistic image of a cat

Validation settings

• CFG: 1.0
• CFG Rescale: 0.0
• Steps: 8
• Sampler: FlowMatchEulerDiscreteScheduler
• Seed: 42
• Resolution: 1024x1024
• Skip-layer guidance: skip_guidance_layers=[],

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

A photo-realistic image of a cat

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: 1

• Training steps: 500

• Learning rate: 0.0001

  • Learning rate schedule: constant_with_warmup
  • Warmup steps: 500

• Max grad value: 0.1

• Effective batch size: 32

  • Micro-batch size: 4
  • Gradient accumulation steps: 1
  • Number of GPUs: 8

• Gradient checkpointing: True

• Prediction type: flow_matching (extra parameters=['flow_schedule_auto_shift', 'shift=0.0'])

• Optimizer: adamw_bf16

• Trainable parameter precision: Pure BF16

• Base model precision: no_change

• Caption dropout probability: 10.0%

• LoRA Rank: 16

• LoRA Alpha: None

• LoRA Dropout: 0.1

• LoRA initialisation style: default

Datasets

photo10k

• Repeats: 0
• Total number of images: ~10040
• Total number of aspect buckets: 2
• 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 = 'bghira/sd35m-reflow'
pipeline = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) # loading directly in bf16
pipeline.load_lora_weights(adapter_id)

prompt = "A photo-realistic image of a cat"
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=8,
    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=1.0,
    skip_guidance_layers=[],
).images[0]

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