Documentation Index
Fetch the complete documentation index at: https://mintlify.com/lllyasviel/Fooocus/llms.txt
Use this file to discover all available pages before exploring further.
Fooocus exposes five performance presets that trade image quality against generation speed. Each preset fixes the number of denoising steps for both text-to-image and upscale/variation (UOV) workflows, and — for the three fastest modes — automatically loads a distillation LoRA that is specifically trained to produce coherent images in very few steps. Choosing the right mode is one of the quickest ways to tune the balance between waiting time and output quality.
| Mode | Display Name | Steps (txt2img) | Steps (UOV) | LoRA File | Restricted Features |
|---|
QUALITY | Quality | 60 | 36 | — | No |
SPEED | Speed | 30 | 18 | — | No |
EXTREME_SPEED | Extreme Speed | 8 | 8 | sdxl_lcm_lora.safetensors | Yes |
LIGHTNING | Lightning | 4 | 4 | sdxl_lightning_4step_lora.safetensors | Yes |
HYPER_SD | Hyper-SD | 4 | 4 | sdxl_hyper_sd_4step_lora.safetensors | Yes |
Extreme Speed, Lightning, and Hyper-SD are restricted modes. Certain advanced features (such as refiner swap and some ControlNet options) are unavailable when these modes are active, because the distillation LoRAs are designed for unconditional fast sampling and do not interact well with those pipelines.
Quality
The highest-fidelity option. Runs 60 steps for standard generation and 36 steps for upscale/variation tasks. No extra LoRA is loaded. Best for final renders where generation time is not a concern.
Speed
The default mode in most presets. Runs 30 steps (txt2img) and 18 steps (UOV). No extra LoRA is loaded. A practical balance between quality and throughput for everyday use.
Extreme Speed (LCM)
Uses the LCM (Latent Consistency Model) distillation LoRA (sdxl_lcm_lora.safetensors) to produce images in just 8 steps. The LCM LoRA is automatically downloaded from Hugging Face on first use. Suitable for rapid prototyping and iterating on composition.
Lightning
Uses the SDXL Lightning 4-step distillation LoRA (sdxl_lightning_4step_lora.safetensors). Generates in only 4 steps. Lightning-distilled models are trained with a different objective than LCM and tend to produce sharper but sometimes less coherent outputs at very low CFG scales.
Hyper-SD
Uses the Hyper-SD 4-step LoRA (sdxl_hyper_sd_4step_lora.safetensors). Also 4 steps. Hyper-SD focuses on high-fidelity preservation at reduced step counts and can produce results competitive with full-step sampling.
Samplers
Fooocus exposes the full ComfyUI k-sampler family plus two extra samplers. The sampler controls the numerical solver used to integrate the diffusion ODE during denoising.
K-Samplers
These samplers correspond to the standard ComfyUI/k-diffusion sampler names. The table shows the Fooocus internal key and the equivalent Civitai / A1111 display name where available.
| Fooocus Key | Civitai / A1111 Name |
|---|
euler | Euler |
euler_ancestral | Euler a |
heun | Heun |
heunpp2 | — |
dpm_2 | DPM2 |
dpm_2_ancestral | DPM2 a |
lms | LMS |
dpm_fast | DPM fast |
dpm_adaptive | DPM adaptive |
dpmpp_2s_ancestral | DPM++ 2S a |
dpmpp_sde | DPM++ SDE |
dpmpp_sde_gpu | DPM++ SDE |
dpmpp_2m | DPM++ 2M |
dpmpp_2m_sde | DPM++ 2M SDE |
dpmpp_2m_sde_gpu | DPM++ 2M SDE |
dpmpp_3m_sde | — |
dpmpp_3m_sde_gpu | — |
ddpm | — |
lcm | LCM |
tcd | TCD |
restart | Restart |
| Fooocus Key | Civitai / A1111 Name |
|---|
ddim | DDIM |
uni_pc | UniPC |
uni_pc_bh2 | — |
The DPM++ family (dpmpp_2m, dpmpp_2m_sde, dpmpp_2m_sde_gpu, etc.) is particularly well-suited for SDXL. SDXL can occasionally produce overly smooth textures, and DPM++ samplers tend to introduce fine-grained detail that counterbalances this, resulting in output that reads as natural to human perception. The default sampler in most Fooocus presets is dpmpp_2m_sde_gpu.
Schedulers
The scheduler determines how noise levels are spaced across the denoising steps. Fooocus supports the following scheduler names:
| Scheduler | Notes |
|---|
normal | Standard linear noise schedule |
karras | Karras et al. noise schedule — default in most presets |
exponential | Exponential noise spacing |
sgm_uniform | SGM uniform spacing |
simple | Simplified schedule |
ddim_uniform | DDIM-style uniform schedule |
lcm | LCM-specific schedule, used with Extreme Speed mode |
turbo | Turbo schedule for fast models |
align_your_steps | Align Your Steps optimized schedule |
tcd | TCD schedule |
edm_playground_v2.5 | EDM schedule tuned for Playground v2.5 |
Civitai Sampler Compatibility
When importing prompts or generation parameters from Civitai, be aware that some samplers do not use the Karras noise schedule even when the karras scheduler is selected. The following samplers bypass Karras regardless of the scheduler setting:
euler, euler_ancestral, heun, dpm_fast, dpm_adaptive, ddim, uni_pc
