Skip to main content
MaxDiffusion supports loading multiple LoRA adapters simultaneously, allowing you to combine different styles and effects in a single generation. This feature is available for Flux and SDXL models.

​
Multiple Flux LoRAs

You can load multiple LoRA adapters for Flux models by providing arrays in the lora_config parameter.

​
Example

python src/maxdiffusion/generate_flux.py \
  src/maxdiffusion/configs/base_flux_dev.yml \
  jax_cache_dir=/tmp/cache_dir \
  run_name=flux_test \
  output_dir=/tmp/ \
  prompt='A cute corgi lives in a house made out of sushi, anime' \
  num_inference_steps=28 \
  ici_data_parallelism=1 \
  ici_fsdp_parallelism=-1 \
  split_head_dim=True \
  lora_config='{"lora_model_name_or_path" : ["/home/jfacevedo/anime_lora.safetensors", "/home/jfacevedo/amateurphoto-v6-forcu.safetensors"], "weight_name" : ["anime_lora.safetensors","amateurphoto-v6-forcu.safetensors"], "adapter_name" : ["anime","realistic"], "scale": [0.6, 0.6], "from_pt": ["true","true"]}'
In this example:
  • Two LoRAs are loaded: an anime style and a realistic photography style
  • Each LoRA has a scale of 0.6, allowing them to blend together
  • Both are loaded from safetensors format (from_pt: "true")

​
Multiple SDXL LoRAs

SDXL supports loading multiple LoRAs from both local paths and HuggingFace Hub.

​
Example: blueprint and papercut styles

python src/maxdiffusion/generate_sdxl.py \
  src/maxdiffusion/configs/base_xl.yml \
  run_name="test-lora" \
  output_dir=/tmp/tmp/ \
  jax_cache_dir=/tmp/cache_dir/ \
  num_inference_steps=30 \
  do_classifier_free_guidance=True \
  prompt="ultra detailed diagram blueprint of a papercut Sitting MaineCoon cat, wide canvas, ampereart, electrical diagram, bl3uprint, papercut" \
  per_device_batch_size=1 \
  diffusion_scheduler_config='{"_class_name" : "FlaxDDIMScheduler", "timestep_spacing" : "trailing"}' \
  lora_config='{"lora_model_name_or_path" : ["/home/jfacevedo/blueprintify-sd-xl-10.safetensors","TheLastBen/Papercut_SDXL"], "weight_name" : ["/home/jfacevedo/blueprintify-sd-xl-10.safetensors","papercut.safetensors"], "adapter_name" : ["blueprint","papercut"], "scale": [0.8, 0.7], "from_pt": ["true", "true"]}'
This example demonstrates:
  • Loading one LoRA from a local file path (blueprint style)
  • Loading another LoRA from HuggingFace Hub (papercut style)
  • Using different scale values (0.8 and 0.7) for fine-tuned blending
  • Combining both effects in the generation

​
Hyper SDXL LoRA

Hyper-SDXL enables ultra-fast inference in just 2-4 steps using specialized LoRA adapters from ByteDance.

​
Features

  • Generate high-quality images in 2-4 inference steps
  • Compatible with custom base models
  • Supports DDIM scheduler with trailing timestep spacing
  • Can disable classifier-free guidance for 1-2 step generation

​
Example: 2-step generation

python src/maxdiffusion/generate_sdxl.py \
  src/maxdiffusion/configs/base_xl.yml \
  run_name="test-lora" \
  output_dir=/tmp/ \
  jax_cache_dir=/tmp/cache_dir/ \
  num_inference_steps=2 \
  do_classifier_free_guidance=False \
  prompt="a photograph of a cat wearing a hat riding a skateboard in a park." \
  per_device_batch_size=1 \
  pretrained_model_name_or_path="Lykon/AAM_XL_AnimeMix" \
  from_pt=True \
  revision=main \
  diffusion_scheduler_config='{"_class_name" : "FlaxDDIMScheduler", "timestep_spacing" : "trailing"}' \
  lora_config='{"lora_model_name_or_path" : ["ByteDance/Hyper-SD"], "weight_name" : ["Hyper-SDXL-2steps-lora.safetensors"], "adapter_name" : ["hyper-sdxl"], "scale": [0.7], "from_pt": ["true"]}'

​
Key parameters for Hyper-SDXL

  • num_inference_steps: Set to 2 or 4 depending on the LoRA variant
  • do_classifier_free_guidance: Set to False for 1-2 steps, True for 4+ steps
  • diffusion_scheduler_config: Must use FlaxDDIMScheduler with trailing timestep spacing
  • scale: Typically 0.7 for optimal results

​
Configuration tips

​
Balancing multiple LoRAs

  • Start with equal scales (e.g., 0.5-0.7 each) and adjust based on results
  • Lower scales produce more subtle effects
  • Higher scales strengthen the LoRA’s influence but may introduce artifacts
  • Total combined scale doesn’t need to equal 1.0

​
LoRA sources

You can mix and match LoRAs from:
  • Local file paths (safetensors or checkpoint files)
  • HuggingFace Hub repositories
  • Different LoRA formats (standard LoRA, LoCON, etc.)

​
Performance considerations

  • Loading multiple LoRAs increases memory usage slightly
  • Inference speed is generally unaffected as LoRAs are merged during initialization
  • All LoRAs are merged using device-side JIT compilation for efficiency

Build docs developers (and LLMs) love

Get started for freeTalk to us