Qwen 3.5 in ComfyUI: prompts and Pixaroma Nodes tutorial

This tutorial uses Qwen 3.5 inside ComfyUI to expand short text into detailed image prompts and describe reference images. The presenter places the model in the text encoders folder, refreshes node definitions with R, and selects it in the workflow. Although ComfyUI loads it through a CLIP-style node, the speaker explains that it is a vision-language LLM that produces text.

The generated prompt then feeds the text encoder in a Flux Klein workflow. Combining both models requires enough VRAM; the presenter reports using 24 GB, without establishing a minimum. In the demonstration, unchanged inputs reuse the existing prompt. Changing the seed manually produces another prompt, while editing the input text also triggers generation. An image example replaces a bunny with a panda through instructions in the description prompt. The aim is an adjustable description rather than an exact reconstruction.

The Pixaroma Nodes walkthrough covers selective image saving, custom ratios, and a visual crop editor. The resolution node uses multiples of eight and limits custom dimensions to 4K. The Note node supports editable download buttons, model folder paths, colored icons, and grids. The Align Tool adds guides and snapping when moving or resizing nodes, with Shift temporarily bypassing it. A final batch preview example shows how to browse outputs and save only a selected image; the presenter warns that batches need more VRAM.