README.md

---
license: other
base_model: "black-forest-labs/FLUX.1-dev"
tags:
  - flux
  - flux-diffusers
  - text-to-image
  - diffusers
  - simpletuner
  - safe-for-work
  - lora
  - template:sd-lora
  - lycoris
inference: true
widget:
- text: 'unconditional (blank prompt)'
  parameters:
    negative_prompt: 'blurry, cropped, ugly'
  output:
    url: ./assets/image_0_0.png
- text: 'In the style of a Harrison Fisher Girl illustration, a woman in a flowing silk wrap dress checks her rose gold iPhone, while her toddler in a lace-trimmed dress peers over her shoulder, both bathed in warm morning light'
  parameters:
    negative_prompt: 'blurry, cropped, ugly'
  output:
    url: ./assets/image_1_0.png
- text: 'In the style of a Harrison Fisher Girl illustration, three young professionals in tailored outfits gather around a sleek laptop at a marble cafe table, sharing quiet laughter as they review the screen'
  parameters:
    negative_prompt: 'blurry, cropped, ugly'
  output:
    url: ./assets/image_2_0.png
- text: 'In the style of a Harrison Fisher Girl illustration, at a grand piano, she wears a vintage-inspired maxi dress while scrolling through sheet music on an iPad, he leans close in a fitted tuxedo jacket'
  parameters:
    negative_prompt: 'blurry, cropped, ugly'
  output:
    url: ./assets/image_3_0.png
- text: 'In the style of a Harrison Fisher Girl illustration, three friends in neo-Edwardian blazers and high-necked blouses review holographic displays in their Tesla, their smart glasses gleaming'
  parameters:
    negative_prompt: 'blurry, cropped, ugly'
  output:
    url: ./assets/image_4_0.png
- text: 'In the style of a Harrison Fisher Girl illustration, a content creator in a puff-sleeve prairie dress adjusts her ring light, her Gibson Girl-inspired updo catching the warm glow'
  parameters:
    negative_prompt: 'blurry, cropped, ugly'
  output:
    url: ./assets/image_5_0.png
- text: 'In the style of a Harrison Fisher Girl illustration, a woman in a designer athleisure set checks her Apple Watch, her 1910s-inspired messy bun perfectly arranged, as her roomba hums nearby'
  parameters:
    negative_prompt: 'blurry, cropped, ugly'
  output:
    url: ./assets/image_6_0.png
- text: 'In the style of a Harrison Fisher Girl illustration, three friends in modernized Gibson Girl dresses pose with their phones for a mirror selfie, their vintage-styled curls contrasting with neon cocktails'
  parameters:
    negative_prompt: 'blurry, cropped, ugly'
  output:
    url: ./assets/image_7_0.png
- text: 'In the style of a Harrison Fisher Girl illustration, an artist in a flowing house dress adjusts her VR headset, her cat perched regally on a Wacom tablet, soft RGB lights highlighting her Gibson-inspired waves'
  parameters:
    negative_prompt: 'blurry, cropped, ugly'
  output:
    url: ./assets/image_8_0.png
---

# HarrisonFisher-EMA-Flux-LoKr

This is a LyCORIS adapter derived from [black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev).


No validation prompt was used during training.

None



## Validation settings
- CFG: `3.0`
- CFG Rescale: `0.0`
- Steps: `20`
- Sampler: `FlowMatchEulerDiscreteScheduler`
- Seed: `42`
- Resolution: `896x1280`
- Skip-layer guidance: 

Note: The validation settings are not necessarily the same as the [training settings](#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: 8
- Training steps: 1500
- Learning rate: 0.0004
  - Learning rate schedule: polynomial
  - Warmup steps: 100
- Max grad norm: 0.1
- Effective batch size: 3
  - Micro-batch size: 3
  - Gradient accumulation steps: 1
  - Number of GPUs: 1
- Gradient checkpointing: True
- Prediction type: flow-matching (extra parameters=['flux_schedule_auto_shift', 'shift=0.0', 'flux_guidance_mode=constant', 'flux_guidance_value=1.0', 'flux_beta_schedule_alpha=8.0', 'flux_beta_schedule_beta=2.0', 'flow_matching_loss=compatible'])
- Optimizer: adamw_bf16
- Trainable parameter precision: Pure BF16
- Caption dropout probability: 10.0%

- SageAttention: Enabled inference
### LyCORIS Config:
```json
{
    "algo": "lokr",
    "multiplier": 1.0,
    "linear_dim": 10000,
    "linear_alpha": 1,
    "factor": 16,
    "apply_preset": {
        "target_module": [
            "Attention",
            "FeedForward"
        ],
        "module_algo_map": {
            "Attention": {
                "factor": 16
            },
            "FeedForward": {
                "factor": 8
            }
        }
    }
}
```

## Datasets

### harrison-fisher-512
- Repeats: 10
- Total number of images: 12
- Total number of aspect buckets: 2
- Resolution: 0.262144 megapixels
- Cropped: False
- Crop style: None
- Crop aspect: None
- Used for regularisation data: No
### harrison-fisher-1024
- Repeats: 10
- Total number of images: 12
- Total number of aspect buckets: 1
- Resolution: 1.048576 megapixels
- Cropped: False
- Crop style: None
- Crop aspect: None
- Used for regularisation data: No
### harrison-fisher-512-crop
- Repeats: 10
- Total number of images: 12
- Total number of aspect buckets: 1
- Resolution: 0.262144 megapixels
- Cropped: True
- Crop style: random
- Crop aspect: square
- Used for regularisation data: No
### harrison-fisher-1024-crop
- Repeats: 10
- Total number of images: 12
- Total number of aspect buckets: 1
- Resolution: 1.048576 megapixels
- Cropped: True
- Crop style: random
- Crop aspect: square
- Used for regularisation data: No


## Inference


```python
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 = 'black-forest-labs/FLUX.1-dev'
adapter_repo_id = 'davidrd123/HarrisonFisher-EMA-Flux-LoKr'
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 = "An astronaut is riding a horse through the jungles of Thailand."


## 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,
    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=896,
    height=1280,
    guidance_scale=3.0,
).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.