🌊🚒 FlowState Flux Engine

A stock ComfyUI Flux graph needs seven or eight nodes just to render one image: a loader, two CLIP loaders, a VAE loader, a dual text-encode, FluxGuidance, a sampler, a decoder. FlowState Flux Engine collapses that whole row into a single box. You pick a model, type a prompt, hit go. It's the node you reach for when you've built the same Flux sampler graph for the tenth time and just want to generate.

What it does under the hood

Nothing exotic - it's a thin, well-organized wrapper over ComfyUI's own building blocks. It loads your Flux model (as a solo diffusion_models/unets file via UNETLoader, or as a checkpoint, which bundles CLIP and VAE), loads the two text encoders with DualCLIPLoader, applies FluxGuidance at your guidance value, and samples with SamplerCustomAdvanced. Crucially it does the whole thing itself and also passes the loaded model, clip, and vae out, so you can still grab them and wire a detailer or upscaler downstream without re-loading anything.

Inputs that matter

• model_filetype - solo_model vs checkpoint. Pick checkpoint and the weight_dtype, clip_1_name, clip_2_name, and vae_name fields are ignored - the tooltip says so, and it's right: they're inside the checkpoint.
• model_name - the Flux model itself. This is the enum with no choices in the schema because it's populated from whatever's in your diffusion_models, checkpoints, and unets folders.
• guidance - defaults to 3.2. Flux is a flow-matching, guidance-distilled model; this is not SDXL, so your old CFG 7 habits do not apply. Leave it near default and tune in small steps.
• steps (32) and denoise (1.0) - at 1.0 it's pure text-to-image; drop denoise below 1 and feed an image via the optional input_img input for img2img work.
• resolution / orientation / latent_type - the preset list (18 ratios, heavy on 16:9 and 21:9 for production work) or Custom, plus whether to start from empty latents or an input image.
• seed, sampling_algorithm (44 choices), scheduling_algorithm (9) - the usual sampler knobs, in case you want euler over dpmpp_2m.
• lora_model + lora_strength, and sage_attention - the two patching options. The LoRA is applied model-only via LoraLoaderModelOnly; Sage Attention is an optional speed patch that only ever shows a real mode if you have Sage Attention and KJNodes installed.

Outputs

model, clip, vae - the loaded components, for reuse. image (IMAGE) - the decoded result. latent (LATENT) - the same result pre-decode, for chaining another pass. You can honestly ignore all but image.

Install

Part of the FlowState Creator Suite, so the pack install is the whole story:

cd ComfyUI/custom_nodes
git clone https://github.com/flowstateeng/FlowState-Creator-Nodes

then restart, or use ComfyUI Manager and search FlowState-Creator-Nodes. The pack has no pip requirements - it wraps ComfyUI core. What you supply is the model: a Flux model (solo or checkpoint), and if solo, the dual CLIP (clip_1 + clip_2) and a VAE.

Common issues

The sage_attention menu stuck on "disabled" is the top complaint, and it's expected: the suite checks at startup whether Sage Attention and KJNodes are both present and only then exposes the modes. Sage Attention itself is famously fiddly to install because it needs a wheel matching your exact PyTorch/CUDA build - grab the right wheel from the SageAttention releases page rather than plain pip install. And if your output looks overcooked, your guidance is too high - that's the number people actually fiddle with. Everything else in this node is a convenience wrapper; the model selection and the guidance slider are where the real control lives.