Flux Advanced Sampler

A stock Flux sampling chain in ComfyUI is usually a handful of nodes: something to set the guidance value, a sampler, and a VAE decode to turn the result back into pixels. FluxAdvancedSampler does all three in one node - you hand it a model, conditioning, a latent, and a VAE, and an image comes out the other end. No separate guidance node, no separate decode step.

That collapsing is the whole pitch. Flux doesn't use classifier-free guidance the way SD1.5/SDXL do - Dev and Schnell are both guidance-distilled, so there's no real negative prompt, and the "guidance" scale that stands in for it is baked into the conditioning at sample time rather than applied as a second pass. This node bakes that step into the sampler call itself, so you get one box instead of three.

What matters

The required inputs are basically "everything Flux sampling needs":

• model, conditioning, latent_image, vae - the usual four. Wire your Flux checkpoint, your positive prompt encoding, an empty or img2img latent, and the matching VAE.
• sampler_name - the full stock ComfyUI sampler list (44 choices). For Flux specifically, stick to Euler-family samplers; it's a flow-matching model, not the DDPM-style architecture most of that list was tuned against.
• scheduler - 9 choices, and this one actually matters more than people expect. Community testing across dozens of samplers found Karras and Exponential schedules fail universally on flow-matching models like Flux - not just underperform, actually fall apart - because they redistribute denoising effort in a way that only makes sense on a curved DDPM trajectory. Use simple or normal.
• guidance (default 3.5) - Flux's stand-in for CFG scale. 3.5 is the widely-used default, but it's also blamed for Flux's "same face" look; a lot of people run 1.8–2.0 for more varied faces, and push it back up to 4.0–4.5 when a realism LoRA is loaded.
• steps (default 20), denoise (default 1), noise_seed.

One output: IMAGE - the fully decoded result, ready to save.

The mechanism, briefly

Under the hood this is doing what a FluxGuidance + sampler + VAEDecode chain would do manually: apply the guidance value to the conditioning, run the actual denoising loop with your chosen sampler/scheduler/steps/seed, then decode the resulting latent through the VAE you supplied. Bundling it means fewer nodes to keep synced when you're iterating on settings, at the cost of not being able to grab the raw latent out mid-chain (there's no LATENT output here - if you need the pre-decode latent for something else, use the separate sampler/decode nodes instead).

Installing it

Search ComfyUI-TkNodes in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/TensorKaze/ComfyUI-TkNodes
cd ComfyUI-TkNodes
pip install -r requirements.txt

Restart ComfyUI. The node itself is light, but Flux is not: the README calls for a Flux.1 Dev or Schnell checkpoint, 6GB VRAM minimum if you're running a GGUF quant, more like 8–12GB for the fp16/fp8 weights directly.

Common issues

Blurry, muddy output. Wrong scheduler is the classic cause on Flux. If you've got Karras or Exponential selected, switch to simple or normal before touching anything else.

Everyone looks the same. That's the well-documented "Fluxmaxxing" effect from the 3.5 guidance default combined with distillation. Drop guidance toward 1.8–2.0, and skip the usual "beautiful, attractive" prompt words - they push toward the same homogenized look harder.

Out of VRAM. Flux's full stack (transformer + T5 text encoder + VAE) is heavy. If you're not already on a quantized checkpoint (GGUF or fp8), that's the first thing to try before blaming this node.

One honest note on where Flux sits in mid-2026: it's not the newest thing in the room anymore - Z-Image, Qwen-Image and Klein have picked up a lot of the local-generation crowd Flux used to have to itself. But there's still a huge library of Flux LoRAs and checkpoints out there, and this node is just solid, standard Flux sampling with nothing fancy layered on - nothing about it stops working because a newer model exists.