anky-flux-lora-v2

anky-flux-lora-v2

Round-two LoRA for Anky on top of FLUX.1-dev.

This model was trained on a deduplicated + caption-complete dataset from the Anky visual universe and is currently used by anky.app/generate (Flux path).

What is in this repo

• Final dataset archive:
  • training-data/final-training-dataset-for-round-two.tar.gz
• Final production weights:
  • training-runs/anky_flux_lora_v2/weights/anky_flux_lora_v2.safetensors
• Recent checkpoints:
  • training-runs/anky_flux_lora_v2/weights/anky_flux_lora_v2_000005000.safetensors
  • training-runs/anky_flux_lora_v2/weights/anky_flux_lora_v2_000005500.safetensors
  • training-runs/anky_flux_lora_v2/weights/anky_flux_lora_v2_000006000.safetensors
• Training samples:
  • training-runs/anky_flux_lora_v2/samples/...
• Repro metadata:
  • training-runs/anky_flux_lora_v2/meta/...

Training summary (run 002)

• Base model: black-forest-labs/FLUX.1-dev
• Technique: LoRA
• Trigger word: anky
• Rank / alpha: 32 / 16
• Steps: 6500
• Batch size: 1
• Optimizer: adamw8bit
• Dtype: bf16
• Dataset size used: 465 image-caption pairs
• Hardware: RunPod RTX PRO 6000 (96GB)

Prompting tips

• Always include anky in prompt.
• Keep scene + emotion explicit.
• Avoid over-describing anatomy; let LoRA carry identity.

Example prompts

1. anky standing on a wet neon street at night, blue skin, glowing amber eyes, painterly cinematic lighting, soft rain, high detail
2. anky writing alone at dawn by a wooden window, purple swirling hair with gold accents, warm sunlight, intimate portrait, textured brushwork

Diffusers usage (example)

import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    torch_dtype=torch.bfloat16,
).to("cuda")

# Option A: local file path
pipe.load_lora_weights("/path/to/anky_flux_lora_v2.safetensors")

# Option B: from this HF repo path
# pipe.load_lora_weights(
#   "jpfraneto/anky-flux-lora-v2",
#   weight_name="training-runs/anky_flux_lora_v2/weights/anky_flux_lora_v2.safetensors"
# )

pipe.fuse_lora(lora_scale=0.85)
image = pipe(
    prompt="anky walking through a market at sunset, painterly, cinematic",
    num_inference_steps=28,
    guidance_scale=3.5,
    width=1024,
    height=1024,
).images[0]
image.save("anky_v2.png")

Notes

• FLUX.1-dev is gated; make sure your Hugging Face account has accepted its license.
• Training and deployment logs are documented on:
  • https://anky.app/trainings/2026-03-04
  • https://anky.app/trainings/general-instructions