Nodes/comfyui-sdnq/SDNQ Sampler
ComfyUI Node

SDNQ Sampler

Load SDNQ quantized models and generate images with 50-75% VRAM savings. Supports FLUX, SD3, SDXL, video models, and more.

By EnragedAntelope·Created 9 months ago·Updated 4 months ago· 82
SDNQ Sampler
  • image1
  • image2
  • image3
  • image4
  • image
  • latent
◄model_selectionFLUX.1-dev-qint8►
◄custom_model_path►
◄prompt►
◄negative_promptblurry, low quality, distorted, deformed, ugly, bad anatomy, bad hands, text, watermark, signature►
◄steps25►
◄cfg7.0►
◄width1024►
◄height1024►
◄seed0►
◄schedulerDPMSolverMultistepScheduler►
◄dtypebfloat16►
◄memory_modebalanced►
◄auto_downloadtrue►
◄lora_selection[None]►
◄lora_custom_path►
◄lora_strength1.00►
◄lora2_selection[None]►
◄lora2_custom_path►
◄lora2_strength1.00►
◄lora3_selection[None]►
◄lora3_custom_path►
◄lora3_strength1.00►
◄lora4_selection[None]►
◄lora4_custom_path►
◄lora4_strength1.00►
◄lora5_selection[None]►
◄lora5_custom_path►
◄lora5_strength1.00►
◄use_quantized_matmultrue►
◄use_xformersfalse►
◄enable_vae_tilingfalse►
◄image_resizeNo Resize►
◄num_frames1►
Categorysampling/SDNQ

Inputs (37)

NameTypeDefaultDescription
model_selectionCOMBOFLUX.1-dev-qint8Select a pre-configured SDNQ model (auto-downloads from HuggingFace) or choose [Custom Path] to specify a local model directory
custom_model_pathSTRINGLocal path to SDNQ model directory (only used when [Custom Path] is selected). Example: /path/to/model or C:\path\to\model
promptSTRINGText description of the image to generate. Be descriptive for best results.
negative_promptSTRINGblurry, low quality, distorted, deformed, ugly, bad anatomy, bad hands, text, watermark, signatureWhat to avoid in the image. Default includes common quality issues. Clear this for no negative prompt. Note: FLUX-schnell (cfg=0) ignores negative prompts.
stepsINT251–150Number of denoising steps. More steps = better quality but slower. 20-30 is typical for most models.
cfgFLOAT7.00–30Guidance scale - how closely to follow the prompt. Higher = more literal. FLUX-schnell uses 0.0, FLUX-dev uses 3.5-7.0, SDXL uses 7.0-9.0.
widthINT102464–4096Image width in pixels. Must be multiple of 8. Larger = more VRAM usage. 1024 is standard for FLUX/SDXL.
heightINT102464–4096Image height in pixels. Must be multiple of 8. Larger = more VRAM usage. 1024 is standard for FLUX/SDXL.
seedINT00–18446744073709550000Random seed for reproducible generation. Same seed + settings = same image.
schedulerCOMBODPMSolverMultistepScheduler⚠️ IMPORTANT: Use FlowMatchEulerDiscreteScheduler for FLUX/SD3/Qwen/Z-Image. Use DPMSolver/Euler/UniPC for SDXL/SD1.5. Wrong scheduler = broken images!
dtypeCOMBObfloat16Model precision. bfloat16 recommended for FLUX (best quality/speed). float16 for older GPUs. float32 for CPU.
memory_modeCOMBObalancedMemory management: 'gpu' = All on GPU (fastest, needs 24GB+ VRAM). 'balanced' = Model offloading (12-16GB VRAM). 'lowvram' = Sequential offloading (8GB VRAM, slowest).
auto_downloadBOOLEANtrueAutomatically download model from HuggingFace if not found locally. Disable to only use local models.
lora_selectionoptCOMBO[None]Select LoRA from ComfyUI loras folder ([None] = disabled, [Custom Path] = use custom path below). LoRAs add styles or concepts to generation.
lora_custom_pathoptSTRINGCustom LoRA path or HuggingFace repo ID (only used when [Custom Path] is selected). Example: /path/to/lora.safetensors or username/lora-repo
lora_strengthoptFLOAT1.00-5–5LoRA 1 influence strength. 1.0 = full strength, 0.5 = half strength, 0.0 = disabled. Negative values invert the LoRA effect. Range: -5.0 to +5.0.
lora2_selectionoptCOMBO[None]Optional second LoRA. Select from ComfyUI loras folder or use [Custom Path].
lora2_custom_pathoptSTRINGCustom path for LoRA 2 (only used when [Custom Path] is selected).
lora2_strengthoptFLOAT1.00-5–5LoRA 2 influence strength.
lora3_selectionoptCOMBO[None]Optional third LoRA. Select from ComfyUI loras folder or use [Custom Path].
lora3_custom_pathoptSTRINGCustom path for LoRA 3 (only used when [Custom Path] is selected).
lora3_strengthoptFLOAT1.00-5–5LoRA 3 influence strength.
lora4_selectionoptCOMBO[None]Optional fourth LoRA. Select from ComfyUI loras folder or use [Custom Path].
lora4_custom_pathoptSTRINGCustom path for LoRA 4 (only used when [Custom Path] is selected).
lora4_strengthoptFLOAT1.00-5–5LoRA 4 influence strength.
lora5_selectionoptCOMBO[None]Optional fifth LoRA. Select from ComfyUI loras folder or use [Custom Path].
lora5_custom_pathoptSTRINGCustom path for LoRA 5 (only used when [Custom Path] is selected).
lora5_strengthoptFLOAT1.00-5–5LoRA 5 influence strength.
use_quantized_matmuloptBOOLEANtrueEnable Triton quantized matmul for 30-80% speedup. Uses torch.compile internally for dequantization. Linux: 'pip install triton'. Windows: 'pip install triton-windows'. Auto-disabled if unavailable.
use_xformersoptBOOLEANfalseEnable xFormers memory-efficient attention for 10-45% speedup. Works with all memory modes (gpu/balanced/lowvram). Auto-fallback to SDPA if xformers not installed or incompatible. Requires: pip install xformers
enable_vae_tilingoptBOOLEANfalseEnable VAE tiling for very large images (>1536px). Prevents out-of-memory errors on high resolutions. Minimal performance impact. Recommended for images >1536x1536.
image1optIMAGEOptional source image for image editing models (Qwen-Image-Edit, ChronoEdit, etc.). Leave unconnected for text-to-image generation.
image2optIMAGEOptional second image for multi-image editing models (Qwen-Image-Edit-2509/2511). Not all models support multiple images.
image3optIMAGEOptional third image for multi-image editing models.
image4optIMAGEOptional fourth image for multi-image editing models.
image_resizeoptCOMBONo ResizeResize input images before processing. Smaller = faster inference, less VRAM. 'No Resize' keeps original dimensions.
num_framesoptINT11–257Number of frames for video models (LTX-2, Wan). Set to 1 for image models. Video models: 161 frames ~= 6 sec at 25fps. Must be divisible by 8+1 for some models.

Outputs (2)

NameTypeDescription
imageIMAGE—
latentLATENT—