v0.6.0 - Aquiles-Image

Release Notes - Aquiles-Image v0.6.0

Compatibility with new models

Image Models

• black-forest-labs/FLUX.2-klein-9b-kv

Video Models

• Lightricks/LTX-2.3

Bug fixes and CLI improvements

• There were corrections to the imports that were overlooked when the pipelines were modularized.
• The CLI now uses Typer instead of Click because Typer offers better parameter handling in commands and provides more atomicity to the configuration of Aquiles-Image

LoRA Load Support (Image Models Only)

Aquiles-Image now supports loading LoRA adapters for image models, both in single-device and distributed inference mode.

A LoRA config is defined via a simple JSON file:

{
  "repo_id": "brushpenbob/Flux-retro-Disney-v2",
  "weight_name": "Flux_retro_Disney_v2.safetensors",
  "adapter_name": "flux-retro-disney-v2",
  "scale": 1.0
}

You can also generate it programmatically:

from aquilesimage.utils import save_lora_config
from aquilesimage.models import LoRAConfig

save_lora_config(
    LoRAConfig(
        repo_id="brushpenbob/Flux-retro-Disney-v2",
        weight_name="Flux_retro_Disney_v2.safetensors",
        adapter_name="flux-retro-disney-v2"
    ),
    "./lora_config.json"
)

Then pass it at startup via the new CLI flags:

aquiles-image serve \
  --model "black-forest-labs/FLUX.1-dev" \
  --load-lora \
  --lora-config "./lora_config.json"

The LoRA is loaded before optimizations (fuse_qkv_projections, FlashAttention) and before torch.compile when enabled, which is the correct order according to the Diffusers documentation.

Installation

uv pip install --upgrade aquiles-image

or

uv pip install aquiles-image==0.6.0

Issues

• bug on main branch #10

• Add support for LTX-2.3 #11

• Add compatibility to model FLUX.2-klein-9b-kv #12

• Add support for loading LoRA from a .json file (Image Models) #14

Pull Rquests

• Feature/add compatibility flux.2 klein 9b kv #13
• PR - Add support for loading LoRA from a .json file (Image Models) - #15