Flux2 SDNQ Sampler V2

If you're running a quantized FLUX.2 model through this pack, this is the sampler you want. The sibling node, SDNQ Sampler V2, works with a broad family of models; this one is a specialized implementation for Flux2's Flow Matching architecture, and the difference shows up most in image-to-image. FLUX.2-dev is 32B of transformer - nobody runs that without quantization and a clear-eyed workflow - so the node that actually drives it deserves the attention.

The flow-matching detail that matters: the Flux2 pipeline doesn't denoise the way a classic diffusion model does, and a generic sampler fudges img2img by passing a strength value and hoping. This node instead initializes the latents from your input image via the pipeline's _encode_vae_image() and uses the sigma schedule itself to control denoise - which means it keeps the full step count and adjusts the starting noise level accurately instead of silently dropping steps. It also patches the VAE decode to force float32 input, dodging the bfloat16/float bias mismatch that crops up with the Flux2 VAE. That's the kind of plumbing you don't notice until the alternative produces a mushy, color-shifted image.

Inputs that matter

• model - from SDNQ Model Loader (or SDNQ LoRA Loader after you've stacked LoRAs). Must be a Flux2 pipeline; this node isn't a general sampler.
• prompt - the actual generation text.
• latent_image - leave empty for text-to-image; feed it from SDNQ VAE Encode for img2img. Width and height come from this tensor, and a 1.0.2 fix means it now respects your input image's size instead of forcing 1024×1024.
• denoise - 0.0 to 1.0, default 1.0. This is your i2i dial: 0.3–0.6 gives you "change the composition, keep the subject" territory; 1.0 is a full redraw.
• steps, cfg, seed - the usual suspects. The tooltip's guidance on cfg is worth trusting: FLUX-dev runs 3.5–7.0. Seed 0 randomizes each run; set it ≥1 to reproduce.
• scheduler - exactly one choice, FlowMatchEulerDiscreteScheduler, because that's the only scheduler the Flux2 pipeline supports. The README's warning applies elsewhere in this pack, but here you literally can't pick wrong.

Notice what's not there: no negative prompt input. Flux2 handles guidance through cfg, so don't go hunting for the negative field - it lives on the general SDNQ Sampler V2 instead.

Output and workflow

Output is a single IMAGE, straight into SaveImage/Preview. The recommended chain: SDNQ Model Loader → (SDNQ LoRA Loader) → Flux2 SDNQ TorchCompile → Flux2 SDNQ Sampler V2, with SDNQ VAE Encode feeding latent_image for i2i.

Install

Same pack, same story:

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

Restart, then find it under sampling/SDNQ/Flux2. Manager works too, but if you hit the security-level block, lower ComfyUI's Security Level to Normal/Disabled or install via the Git URL above.

Troubleshooting

OOM at high resolution is the most common complaint - enable_vae_tiling lives on the Model Loader, so flip it there for anything >1536px, or drop to lowvram mode. If i2i comes out looking like the model ignored your image, your denoise is likely too high or the latent_image isn't coming from SDNQ VAE Encode. And if results look broken or uniform, check the console: a wrong scheduler choice would do it, but since this node only exposes the one, the usual culprit is the model not being a real Flux2 pipeline at all.