⭐ Star SD(XL) Start(t) Settings

The classic ComfyUI SDXL graph starts with CheckpointLoader, a CLIPTextEncode for the positive and another for the negative, an EmptyLatentImage, and a seed - five nodes before you've even thought about sampling. SD(XL) Star(t) Settings (SDXLStartSettings) folds that whole opening into one node: load the checkpoint, load a VAE, encode your positive and negative text, build the empty latent, and hand you model, clip, vae, latent, width, height, conditioning_POS, and conditioning_NEG - everything a SDstarsampler needs on a platter.

How it works

Under the hood it's exactly what you'd wire by hand: load_checkpoint_guess_config for the checkpoint (with a VAE output and a CLIP), then it encodes text and negative_text with the CLIP (including the pooled output SDXL uses), builds a zero latent from your chosen ratio, and applies any LoRAs from a LoRA_Stack to both the model and the conditioning CLIP before returning. One detail worth respecting: it encodes the negative for you, which is easy to overlook in hand-built SDXL graphs and quietly changes results when missing.

The inputs that matter

• text and negative_text - your two prompts. Multiline, no surprises.
• Checkpoint - dropdown of your checkpoints. This is the one input you'll actually switch constantly as you A/B models.
• VAE - "Default" (use the checkpoint's VAE) or pick a separate one. Leave it on Default unless you have a tuned VAE you trust more.
• Latent_Ratio - the preset dropdown: 1:1 at 1024, plus landscape and portrait ratios at SDXL-friendly resolutions. This is where you set composition before you generate. Latent_Width/Latent_Height (step 16) only matter in "Free Ratio" mode.
• Batch_Size - how many images per run.
• LoRA_Stack (optional) - feed the lora_stack output from Star 3 LoRAs in here, and the starter applies them internally to model and clip before conditioning.

Outputs: the eight listed above. WIDTH/HEIGHT as ints are handy for downstream size math; conditioning_POS/conditioning_NEG go to the sampler's positive/negative.

The honest tradeoff

The pitch is "one node instead of five," and that's real - shared workflows get dramatically more readable. The tradeoff is that you lose the explicit control of a hand-wired graph: no separate VAE loader node to tap mid-graph, no obvious place to insert a token counter, and if you need to apply a LoRA only to the CLIP for conditioning but not the model, the internal application path is less flexible than doing it yourself. For 95% of SDXL work - especially template workflows you hand to people - the starter wins. For the other 5%, keep the hand-built graph around.

Install

Standard pack install:

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

or ComfyUI Manager → search Starnodes, restart, find it under ⭐StarNodes/Starters. No model downloads - it loads whatever checkpoints you already have. The Checkpoint dropdown being empty means your models/checkpoints folder is empty, not that the node is broken.

Wire it: SDXLStartSettings → SDstarsampler → whatever you're using for preview/save. For a hires-fix, you can insert the Star Model Latent Upscaler between the sampler's latent output and a second sampler at low denoise.