TheNoise is fully integrated with Lemonade Server and ships as its thenoise backend. You can generate and edit images directly from the Lemonade web UI or through its HTTP API.
Ready-made Lemonade recipes for every supported model - checkpoints, labels
and generation defaults — are in recipes/; each model page
links to its own recipes.
Where available, pick a model's INT8-ConvRot recipe over the BF16 one: its weights are quantized to INT8 with a rotation-based scheme, so the download and memory footprint shrink substantially and generation runs faster, while quality holds up well.
A chosen recipe can be imported via the CLI using the command below.
lemonade import "path/to/recipe.json"Alternatively, it can also be imported from the Lemonade Server web UI: download a recipe you want to try, go to
http://localhost:13305/, choose File → New Model → From JSON, select the recipe file, and wait for the model to finish downloading.
A supported model with a custom checkpoint can also be registered and pulled with a single command (example: Anima Turbo, matching recipes/anima/Anima-Turbo.json):
lemonade pull user.Anima-Turbo \
--checkpoint main circlestone-labs/Anima:split_files/diffusion_models/anima-turbo-v1.1.safetensors \
--checkpoint text_encoder circlestone-labs/Anima:split_files/text_encoders/qwen_3_06b_base.safetensors \
--checkpoint vae circlestone-labs/Anima:split_files/vae/qwen_image_vae.safetensors \
--recipe thenoiseTo use the latest TheNoise version, update your Lemonade config. The lemonade config set command requires administrative privileges or an Admin API key because it modifies internal configuration endpoints. The key can be set explicitly via the LEMONADE_API_KEY environment variable before running the command.
lemonade config set thenoise.rocm_bin="latest"
# create loras dir
sudo mkdir -p /var/cache/thenoise/loras
# grant permissions to lemonade
sudo chmod -R lemonade:lemonade /var/cache/thenoise/loras
# configure lora dir
lemonade config set thenoise.lora_dir="/var/cache/thenoise/loras"Example configuration for Arch Linux.
To use a LoRA with a model:
- Download the LoRA you want and copy its
.safetensorsfile into the configured LoRA directory. - In the model's recipe
.jsonfile, add a"lora_specs"entry to the"recipe_options"block, with the LoRA name and its weight (multiple LoRAs are comma-separated):"lora_specs": "lora-name:0.8". - Import the updated recipe in the Lemonade Server web UI at
http://localhost:13305/(File → New Model → From JSON).