wan2.1-1.3b-anyflow-wip

bghira/wan2.1-1.3b-anyflow-wip

This is a PEFT LoRA derived from Wan-AI/Wan2.1-T2V-1.3B-Diffusers.

The main validation prompt used during training was:

A vibrant green Mustang GT parked in an empty parking lot. The camera slowly pans around the car, showing its sleek design, black hood and black rims in a clean promotional video.

Validation settings

• CFG: 1.0
• CFG Rescale: 0.0
• Steps: 4
• Sampler: AnyFlowValidationScheduler (FlowMatchEulerDiscreteScheduler)
• Seed: 0
• Resolution: 832x480

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

You can find some example images and videos 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: 6000

• Learning rate: 6e-05

  • Learning rate schedule: constant_with_warmup
  • Warmup steps: 1000

• Max grad norm: 1.0

• Effective batch size: 8

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

• Gradient checkpointing: False

• Prediction type: flow_matching (extra parameters=['shift=5.0'])

• Optimizer: torch-adamw (config=weight_decay=0.0,eps=1e-8)

• Trainable parameter precision: Pure BF16

• Base model precision: no_change

• Caption dropout probability: 0.1%

• LoRA Rank: 256

• LoRA Alpha: 256.0

• LoRA Dropout: 0.0

• LoRA initialisation style: default

• LoRA mode: Standard

Datasets

wan21-anyflow-openvid-81f-832x480

• Repeats: 0
• Total number of videos: 8192
• Total number of aspect buckets: 1
• Target frame count: 81
• FPS: 16
• Used for regularisation data: No

Inference

import torch
from diffusers import DiffusionPipeline

model_id = 'Wan-AI/Wan2.1-T2V-1.3B-Diffusers'
adapter_id = 'bghira/wan2.1-1.3b-anyflow-wip'
pipeline = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) # loading directly in bf16
pipeline.load_lora_weights(adapter_id)

prompt = "A vibrant green Mustang GT parked in an empty parking lot. The camera slowly pans around the car, showing its sleek design, black hood and black rims in a clean promotional video."
negative_prompt = '色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走'

## 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=4,
    generator=torch.Generator(device='cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu').manual_seed(1),
    width=832,
    height=480,
    guidance_scale=1.0,
).images[0]

from diffusers.utils.export_utils import export_to_gif
export_to_gif(model_output, "output.gif", fps=16)

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.