anima-anyflow-e621-dmd-v4

ComfyUI support: SimpleTuner-io/ComfyUI-AnyFlow is the canonical AnyFlow ComfyUI integration and ships with a ready-made workflow for Anima few-step inference.

Experiment: AnyFlow on-policy DMD, v4 (EMA ablation: 5-step interval + fast decay)

Fourth cell of the stage-2 EMA ablation. Identical to dmd-v3 — same stage-1 EMA init, objective, and data — except ema_update_interval: 5 with ema_decay: 0.99, i.e. v1's sparse cadence combined with v3's fast decay (~300 EMA updates over the 1,500-step run).

Outcome: confirms update cadence, not decay, is the binding constraint at this run length — the sparsely-updated EMA lags both v2 and v3 regardless of the shorter window. Grid conclusion: per-step updates are required for a useful stage-2 EMA; decay is secondary. Live weights remain the ship candidate.

---

bghira/anima-anyflow-e621-dmd-v4-wip

This is a PEFT LoRA derived from circlestone-labs/Anima-Base-v1.0-Diffusers.

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: 1.0
• CFG Rescale: 0.0
• Steps: 4
• Sampler: AnyFlowValidationScheduler (FlowMatchEulerDiscreteScheduler)
• Seed: 42
• Resolution: 1024x1024

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: constant_with_warmup
  • Warmup steps: 250

• Max grad value: 1.0

• Effective batch size: 16

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

• Gradient checkpointing: True

• Prediction type: flow_matching[]

• Optimizer: adamw_bf16

• Trainable parameter precision: Pure BF16

• Base model precision: no_change

• Caption dropout probability: 0.1%

• LoRA Rank: 128

• LoRA Alpha: 128.0

• LoRA Dropout: 0.1

• LoRA initialisation style: default

• LoRA mode: Standard

Datasets

anima-anyflow-e621-1024

• Repeats: 0
• Total number of images: ~285096
• 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 = 'circlestone-labs/Anima-Base-v1.0-Diffusers'
adapter_id = 'bghira/anima-anyflow-e621-dmd-v4-wip'
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"
negative_prompt = 'worst quality, low quality, score_1, score_2, score_3, blurry, cropped, artist name, signature'

## 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(42),
    width=1024,
    height=1024,
    guidance_scale=1.0,
).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.