bghira/flux-anyflow-e621-dmd

About this experiment

FLUX.1-dev was distilled into an n-step model using AnyFlow (a MeanFlow-style forward stage followed by on-policy DMD) on the e621 furry split. Two deliberate choices shaped the run:

1. 1.Why start from a distilled base? FLUX.1-dev is itself a distilled model. This run was a check for mistakes in the AnyFlow implementation — specifically, whether already-distilled models simply can't be AnyFlowed. They can.
2. 2.Why e621? It was readily available via webshart in SimpleTuner.

The adapter was intentionally not trained long enough to resolve the remaining 2-step and 1-step generation difficulty: the objective above was validated, and training stopped there.

bghira/flux-anyflow-e621-dmd

This is a PEFT LoRA derived from black-forest-labs/flux.1-dev.

The main validation prompt used during training was:

masterpiece, best quality, score_7, safe, anime portrait of a young woman with blue hair wearing a white jacket, clean line art, detailed eyes, soft daylight

Validation settings

• —CFG: 3.5
• —CFG Rescale: 0.0
• —Steps: 4
• —Sampler: AnyFlowValidationScheduler (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:

<Gallery />

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

Training settings

• —Training epochs: 0
• —Training steps: 1500
• —Learning rate: 2e-06
• —Learning rate schedule: constantwithwarmup
• —Warmup steps: 250
• —Max grad value: 1.0
• —Effective batch size: 4
• —Micro-batch size: 1
• —Gradient accumulation steps: 1
• —Number of GPUs: 4
• —Gradient checkpointing: True
• —Prediction type: flowmatching (extra parameters=['shift=3.0', 'fluxguidancemode=constant', 'fluxguidancevalue=3.5', 'fluxlora_target=all'])
• —Optimizer: adamw_bf16
• —Trainable parameter precision: Pure BF16
• —Base model precision: no_change
• —Caption dropout probability: 0.0%

• —LoRA Rank: 128
• —LoRA Alpha: 128.0
• —LoRA Dropout: 0.0
• —LoRA initialisation style: default
• —LoRA mode: Standard

Datasets

flux-e621-1024

• —Repeats: 0
• —Total number of images: 65536
• —Total number of aspect buckets: 1
• —Resolution: 1.048576 megapixels
• —Cropped: True
• —Crop style: random
• —Crop aspect: square
• —Used for regularisation data: No

Inference

import torch
from diffusers import DiffusionPipeline

model_id = 'black-forest-labs/flux.1-dev'
adapter_id = 'bghira/flux-anyflow-e621-dmd'
pipeline = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) # loading directly in bf16
pipeline.load_lora_weights(adapter_id)

prompt = "masterpiece, best quality, score_7, safe, anime portrait of a young woman with blue hair wearing a white jacket, clean line art, detailed eyes, soft daylight"


## 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,
    num_inference_steps=4,
    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=3.5,
).images[0]

model_output.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.