SDNQ Sampler V2

This is the general workhorse of the comfyui-sdnq-splited pack: feed it a quantized model from SDNQ Model Loader, get an image out. It's the fork's answer to the original pack's one-node-does-everything design - model loading got split off into its own loader, and this sampler just sits at the end of the chain like a KSampler you already know. The catch, and it's a real one: the pack is developed and tested for FLUX.2 only. If you're on FLUX.2, use the specialized Flux2 SDNQ Sampler V2 instead, and treat this node as the option for the rest of the SDNQ catalog - FLUX.1, Qwen-Image, SD3-class models, and anything else the loader's dropdown offers.

The inputs that matter

• model - from SDNQ Model Loader, or SDNQ LoRA Loader once you've stacked adapters.
• prompt and negative_prompt - the negative has a sensible default ("blurry, low quality, distorted…") you'll probably just keep. One trap: if you run a schnell-class model at cfg 0, negative prompts are ignored entirely.
• steps, cfg, seed - standard, with the tooltip's numbers worth reading: FLUX/schnell runs at cfg 0, FLUX-dev at 3.5–7.0, SDXL at 7–9.
• scheduler - the one field that will wreck your output if you get it wrong, and the README puts it in caps: wrong scheduler = broken images. The rule of thumb is simple: FlowMatchEulerDiscreteScheduler for FLUX/SD3/Qwen/Z-Image; DPMSolver/Euler/UniPC family for SDXL/SD1.5. The dropdown has 14 entries; pick by model family, not by vibes.
• latent_image - empty for t2i, or from SDNQ VAE Encode for i2i, with denoise (0–1) controlling how much of the source survives.

Output is a single IMAGE, straight into SaveImage/Preview.

Why the Flux2 sampler exists

All the flow-matching detail - accurate denoise via sigma schedule, VAE float32 patching, compute_empirical_mu handling - is in the Flux2 node, not here. This sampler is more generic, which is precisely why it's wrong for your FLUX.2 workflow. It will run a Flux2 model, but the README is explicit that the Flux2-optimized node gives better i2i results and more accurate denoise control. Use this one for the non-Flux2 models in the catalog, and reach for Flux2SDNQSamplerV2 for the thing the pack was actually built for.

Install

Nothing pack-specific here that isn't in the README:

cd ComfyUI/custom_nodes/
git clone https://github.com/ussoewwin/comfyui-sdnq-splited.git
cd comfyui-sdnq-splited
pip install -r requirements.txt

Restart ComfyUI, and the node appears under sampling/SDNQ. Manager install works but may hit the security-level message - lower the Security Level to Normal/Disabled or use "Install via Git URL" with the repo above. Note the dependency weight: sdnq, diffusers>=0.36.0 (may need installing straight from the huggingface/diffusers GitHub repo), transformers, accelerate.

Troubleshooting

The two failures everyone hits: broken/noisy images = wrong scheduler for the model family (see above), and OOM at high resolutions = set enable_vae_tiling on the Model Loader, drop memory_mode a step, or go more aggressive on the quantization (uint4 over int8). If the sampler throws on load, check that the model you selected is genuinely SDNQ-quantized from Disty0's collection - the loader's custom path option will happily point at a folder that isn't one.