⭐ Star Qwen Image Start(t) Settings

A Qwen-Image workflow normally starts with the same five nodes every time: load the diffusion model, load the VAE, load the CLIP, make an empty latent, then CLIP-encode the prompt. Star Qwen Image Start(t) Settings collapses that whole front of the graph into one node that outputs the model, the clip, the vae, the latent and both conditioning - prompt handling included.

It's in the ⭐StarNodes/Starters category, alongside the pack's FLUX and SD(XL) start nodes. Same family, same idea: one node to kick off a whole workflow.

How it works

Give it a positive and negative prompt, and it does the rest internally: loads the diffusion model from models/diffusion_models or models/unet, the VAE from models/vae, and the CLIP from models/clip using CLIP_Type qwen_image (the setting for Qwen models - and it defaults there, so you rarely have to touch it). It creates an empty latent at a Qwen-native resolution, encodes both prompts into CONDITIONING, and hands everything to you on ten outputs.

The resolution handling is the part worth paying attention to. Qwen-Image's official ratio table is baked in as presets - 1:1 (1328×1328), 16:9 (1664×928), 9:16 (928×1664), 4:3, 3:4, 3:2, 2:3, 5:7, 7:5, and a Free Ratio (custom) mode with Latent_Width/Latent_Height. These aren't random sizes; they're the native resolutions Qwen was trained on, and generating off-table costs you quality. Two flags make this friendlier:

• use_nearest_image_ratio - with an image connected, it auto-selects the closest preset ratio to that image. Great for img2img where the latent should match the source.
• LoRA_Stack (optional) - feed a LoRA stack and it's applied to the model and the internal conditioning before anything leaves the node.

The outputs that matter

Ten outputs, but you'll use six on a normal graph: model, clip, vae, latent, condition_pos, condition_neg. The rest - width, height, prompt_pos, prompt_neg - are the plain numbers/strings, useful when a downstream node needs the actual values rather than the objects.

Two notes on defaults: CLIP_Device defaults to cpu, which is deliberate - the Qwen CLIP is a 4B model and offloading it to CPU keeps VRAM for the diffusion model. And Batch_Size just controls how many latents are created (1 is fine unless you're batching).

Installing it

StarNodes installs through ComfyUI Manager (search Starnodes, install, restart) or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

Find it by double-clicking the canvas and typing star - it's under ⭐StarNodes/Starters.

Where people get burned

Two spots. First, the dropdowns say Default until you've got models in the right folders - an empty models/clip means the CLIP selector finds nothing, so make sure your Qwen CLIP is actually in models/clip before you blame the node. Second, Default is not magic: if the node can't resolve a model it falls back rather than erroring loudly, and a workflow that "worked" with placeholder models will silently give you weak results. Set all three selectors explicitly the first time. And if you're coming from an SDXL habit, remember Qwen's latent is 16 channels - this node builds that correctly, but if you swap in an external latent from an SDXL node later, that mismatch will bite.