LLS Qwen Text To Image

Qwen-Image is a real model, but running it in ComfyUI is not one node's job in stock form - you need the diffusion model, the Qwen2.5-VL text encoder, the Qwen VAE, plus the right sampling settings (shift, cfg norm, reference latent method). LLS Qwen Text To Image is the "compressed" version: it loads the Qwen companion resources internally when the node runs, so your graph stays at a single node instead of a five-node Qwen sub-graph.

Pick the model from model_name (populated from the Qwen-Image family files in ComfyUI/models/diffusion_models/), type a prompt, pick a width/height, and hit run. The node handles the CLIP, VAE, and UNet loading, then samples and hands you an image out.

The inputs that matter

• model_name - the Qwen text-to-image checkpoint. The dropdown shows "(no qwen text-to-image models found)" until you put one in diffusion_models.
• prompt, width, height, batch_size - the obvious ones.
• negative_prompt, cfg (default 4), sampler_name (default euler), scheduler (default simple), steps (20) - standard sampling controls, all folded under the "advanced" section.
• shift (default 3.1) and cfg_norm_strength - Qwen-family sampling parameters that don't exist on the SD knobs. Leave shift at 3.1 unless you know why you're changing it.
• enable_turbo_mode + turbo_lora_name / turbo_strength - bolt a step-reduction LoRA on to run Qwen faster.

There's also an optional model input, so you can chain an external MODEL in instead of letting the node load its own - useful if you want a quantized GGUF or a merged checkpoint in the loop.

How it loads things

Under the hood it calls ComfyUI's own Qwen runtime pieces - CLIPLoader with the Qwen image clip type, VAELoader, UNETLoader - and it expects the standard companion filenames (qwen_2.5_vl_7b-family text encoders, qwen_image_vae.safetensors). If your ComfyUI build doesn't include official Qwen image support, the node errors with a clear "use a ComfyUI build that includes official Qwen image support" message.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/Gin3601/LLS-node
# restart ComfyUI

Or ComfyUI Manager → search "LLS-node". No pip deps - but the model is the heavy lift. Qwen-Image is a 20B MMDiT; at bf16 that's out of reach for most consumer cards, so in practice people run community quantizations (GGUF, fp8). Budget accordingly before you wire this into a workflow.

Gotchas

The model dropdown is empty until you supply the files - that's expected, not a bug. And the pack's own Simple Checkpoint Loader does not load Qwen; these Qwen nodes self-load. If the node errors about a missing runtime class, update ComfyUI rather than the pack - the Qwen support landed in core comfy_extras.