Fooocus on Thumper

Installing Fooocus manually requires cloning the repo, setting up Python, downloading SDXL models, and configuring GPU-specific flags. Thumper handles all of this in one click.

Fooocus ships with an SDXL-based model by default. It includes built-in style presets (cinematic, anime, photographic, etc.) that modify the generation pipeline without requiring manual prompt engineering.

Default Model

The default checkpoint is an SDXL variant optimized for Fooocus. Style presets are applied as additional prompt conditioning and LoRA weights — no separate model downloads needed.

Changing the Default Model

To use a different checkpoint, edit the Fooocus config file or place a new .safetensors file in the models directory:

# Model directory:
~/.local/share/tr-desktop/apps/fooocus/models/checkpoints/
 
# Set default model in config:
"default_model": "your-model.safetensors"

• Prompt-only generation — type and click Generate
• Style presets: cinematic, anime, photographic, and more
• Inpaint and outpaint for editing existing images
• Upscaling with built-in upscale models
• Image variation — generate alternatives from a source image
• Describe (reverse prompt) — generate a text prompt from an image

Thumper auto-detects your GPU and applies appropriate launch flags. Here are the vendor-specific details:

NVIDIA (CUDA)

• Works out of the box with CUDA 11.8+
• For 4 GB VRAM cards, --lowvram is applied automatically
• FP16 is the default precision for best speed

AMD Discrete GPU (ROCm)

• Requires ROCm 5.7+
• Set TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 for flash attention
• First run may be slow while MIOpen builds its Find-DB

AMD APU (Integrated Graphics)

• Use --disable-offload-from-vram to prevent partial eviction OOM
• BF16 is mandatory — FP16 causes overflow in SDXL attention layers

Apple Silicon (MPS)

• Set PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0 for maximum memory usage
• Unified memory — no explicit VRAM limit

CPU Only

• Very slow — basic images only, expect 5–10 minutes per generation
• FP32 precision is used automatically

Both tools generate images with SDXL, but they serve very different workflows:

LoRA (Low-Rank Adaptation) files add trained styles or subjects to the base model. To use custom LoRAs with Fooocus:

# 1. Place LoRA files in the models directory:
~/.local/share/tr-desktop/apps/fooocus/models/loras/
 
# 2. Supported formats:
your-style.safetensors    # SDXL-compatible LoRA
 
# 3. Enable in Fooocus UI:
# Advanced > LoRA tab > select your LoRA and set weight (0.5-1.0)