⭐ Star FLUX Start(t) Settings

A bare Flux workflow in ComfyUI starts with five or six nodes doing setup: load the UNet, load the two text encoders, load the VAE, create an empty latent, encode your prompt, maybe apply a LoRA in between. FLUX Star(t) Settings (FluxStartSettings) does all of that in one node and hands you every output you need: UNET, CLIP, LATENT, CONDITIONING, VAE, plus WIDTH and HEIGHT as integers. It's the Flux half of the pack's starter trio (the others cover SDXL and SD3.5), and it's designed to plug straight into Fluxstarsampler.

How it works

Inside, it's doing what you'd otherwise do by hand: loading the model from your models/diffusion_models, the two CLIPs (the T5-XXL and the CLIP-L/ViT-L encoder) from text_encoders, and the VAE from vae. It encodes your text into conditioning using both encoders, applies any LoRAs from a LoRA_Stack to the model and the internal conditioning, then builds an empty latent at your chosen size. Nothing about the math is special - the point is that it's one box instead of five, and the box never forgets the CLIP-2 connection that breaks every hand-built Flux graph.

The inputs that matter

• text - your prompt. That's the whole input side of conditioning.
• UNET, CLIP_1, CLIP_2, VAE - dropdowns of what's installed, defaulting to the standard Flux filenames (flux1-dev.safetensors, t5xxl_fp16.safetensors, the ViT-L encoder, ae.safetensors). If you run GGUF or fp8 Flux, pick your files here.
• Weight_Dtype - fp8 options (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2) for low-VRAM setups. Default is default; only reach for fp8 if VRAM is actually the constraint.
• Latent_Ratio - the pack's dropdown of preset aspect ratios (1:1 at 1024, 16:9 at 1344x768, etc.), all at Flux-friendly resolutions. This is the input you'll actually change most.
• Latent_Width / Latent_Height - used when the ratio is "Free Ratio". Both step in 64s and cap at 8192, keeping you roughly on the divisible-by-64 grid Flux likes.
• Batch_Size - up to 64, but each image costs VRAM.
• CLIP_Device (cpu by default - text encoding is fine on CPU and keeps VRAM free) and VAE_Device (cuda) are sensible defaults; leave them alone.
• LoRA_Stack - optional, and this is the integration point: the lora_stack output from the pack's Star 3 LoRAs plugs in here, and the starter applies it internally to both model and conditioning.

Outputs: wire UNET, CLIP, LATENT, CONDITIONING and VAE into a Fluxstarsampler. The WIDTH/HEIGHT ints are handy for nodes that need the resolution as numbers (text overlays, size calculations) without parsing the latent.

Install

Pack install, as always:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes

or ComfyUI Manager → search Starnodes. Restart, find it under ⭐StarNodes/Starters. The model files are yours - this node loads whatever you already have for Flux, it doesn't fetch anything.

Gotchas

The dropdown defaults are names, not downloads - if flux1-dev.safetensors isn't in models/diffusion_models, the node errors on load. If you're using a quantized Flux, change both the UNET and CLIP_1/CLIP_2 to match what you actually have. And if your graph feels like it lost the CLIP connection, remember this node encodes conditioning internally - you don't need a separate CLIPTextEncode, and connecting one can fight the starter's own output.