sd35-terminus-relay-v2

Stable Diffusion 3.5 Medium: Terminus Relay (v2)

This is a LyCORIS adapter derived from stabilityai/stable-diffusion-3.5-medium.

For the earlier first release of Terminus Relay, see terminusresearch/sd35-terminus-relay-v1.

The main validation prompt used during training was:

A photo-realistic image of a cat

Validation settings

• CFG: 3.2
• CFG Rescale: 0.0
• Steps: 30
• Sampler: FlowMatchEulerDiscreteScheduler
• Seed: 42
• Resolution: 1024x1024
• Skip-layer guidance: skip_guidance_layers=[7, 8, 9],

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

You can find some example images in the following gallery:

<Gallery />

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

Training settings

• Training epochs: 2
• Training steps: 295001
• Learning rate: 5e-05
  • Learning rate schedule: constant_with_warmup
  • Warmup steps: 10000
• Max grad norm: 0.1
• Effective batch size: 12
  • Micro-batch size: 4
  • Gradient accumulation steps: 1
  • Number of GPUs: 3
• Gradient checkpointing: True
• Prediction type: flow-matching (extra parameters=['shift=3.0'])
• Optimizer: bnb-adamw8bit
• Trainable parameter precision: Pure BF16
• Caption dropout probability: 10.0%

LyCORIS Config:

{
    "bypass_mode": true,
    "algo": "lokr",
    "multiplier": 1.0,
    "full_matrix": true,
    "linear_dim": 10000,
    "linear_alpha": 1,
    "factor": 4,
    "apply_preset": {
        "target_module": [
            "Attention",
            "FeedForward"
        ],
        "module_algo_map": {
            "FeedForward": {
                "factor": 4
            },
            "Attention": {
                "factor": 2
            }
        }
    }
}

Inference

import torch
from diffusers import DiffusionPipeline
from lycoris import create_lycoris_from_weights


def download_adapter(repo_id: str):
    import os
    from huggingface_hub import hf_hub_download
    adapter_filename = "pytorch_lora_weights.safetensors"
    cache_dir = os.environ.get('HF_PATH', os.path.expanduser('~/.cache/huggingface/hub/models'))
    cleaned_adapter_path = repo_id.replace("/", "_").replace("\\", "_").replace(":", "_")
    path_to_adapter = os.path.join(cache_dir, cleaned_adapter_path)
    path_to_adapter_file = os.path.join(path_to_adapter, adapter_filename)
    os.makedirs(path_to_adapter, exist_ok=True)
    hf_hub_download(
        repo_id=repo_id, filename=adapter_filename, local_dir=path_to_adapter
    )

    return path_to_adapter_file
    
model_id = 'stabilityai/stable-diffusion-3.5-medium'
adapter_repo_id = 'bghira/sd35m-photo-1mp'
adapter_filename = 'pytorch_lora_weights.safetensors'
adapter_file_path = download_adapter(repo_id=adapter_repo_id)
pipeline = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) # loading directly in bf16
lora_scale = 1.0
wrapper, _ = create_lycoris_from_weights(lora_scale, adapter_file_path, pipeline.transformer)
wrapper.merge_to()

prompt = "A photo-realistic image of a cat"
negative_prompt = 'ugly, cropped, blurry, low-quality, mediocre average'

## Optional: quantise the model to save on vram.
## Note: The model was quantised during training, and so it is recommended to do the same 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
image = pipeline(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=30,
    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=3.2,
    skip_guidance_layers=[7, 8, 9],
).images[0]
image.save("output.png", format="PNG")

Exponential Moving Average (EMA)

SimpleTuner generates a safetensors variant of the EMA weights and a pt file.

The safetensors file is intended to be used for inference, and the pt file is for continuing finetuning.

The EMA model may provide a more well-rounded result, but typically will feel undertrained compared to the full model as it is a running decayed average of the model weights.