Qwen TeleStyle Apply

This is the node that does the actual work in the Qwen-TeleStyle pack. The Loader up front is just the shopping trip - it grabs the weights and assembles the pipeline. Apply is where your content photo meets someone else's style, and the "keep Figure 1's content, take Figure 2's look" promise actually gets kept.

What it does

You connect the models socket from the Loader, drop a content_ref and a style_ref image in, type a prompt, hit queue. Out comes a stylized image, plus - unusually - three text outputs. The whole TeleStyleV2 inference routine is reproduced line-for-line from the original reference app's infer(), so what you get locally is meant to match the official reference implementation, not a loose approximation.

How it works

The part worth understanding is that this is not plain two-image img2img. The reference images feed Qwen2.5-VL - the 7B text encoder this pack loads - and if you flip on infer_content_prompt or infer_style_prompt, that vision model writes its own short caption of each reference: a few comma-separated words about what's in the content image, or about five words describing the style. Those captions get merged into your prompt before generation. That's the "tele" in TeleStyle - style and content are named in text, not just implied by pixels.

Your refs get resized so the long edge lands on resolution (rounded down to a multiple of 16, aspect ratio preserved), and the merged prompt runs through the pipeline at the settings you pick.

The inputs that matter

• content_ref / style_ref - both are technically optional, but the node raises an error if you connect neither. Content is the subject you keep; style is the look you borrow.
• steps - defaults to 4, and that's not a typo. The Loader fuses a Lightning step-distillation LoRA into the pipeline, so four steps is the intended speed, not a budget compromise.
• cfg - defaults to 1.0, and leave it there. Qwen-Image-Edit is guidance-distilled; at CFG 1 there's no unconditional pass for a negative prompt to steer, so don't bother wiring one in.
• resolution - the long edge of the output image.
• infer_content_prompt / infer_style_prompt - the auto-caption toggles above. They need the text encoder file with its lm_head included; the loader's copy has it, and you get an explicit error instead of a silent wrong result if you ever hit one without it.

The outputs

image is the stylized result. content_prompt, style_prompt and prompt are strings - the generated captions and the final merged prompt. ComfyUI has no Gradio-style auto-filled textbox, so the pack exposes them as sockets: wire them into a Show Text node to actually read them.

Troubleshooting

Qwen-Image-Edit re-emits the whole frame, it doesn't patch just the region you pointed at. Pixels nobody asked to change can drift, and faces are the usual victim - that's a property of the base model, not a bug in this node, and inpainting the face back is the standard community fix. People also report TeleStyle output that "always looks like it's not fully resolving" - softer than they'd like. If yours does, that's your cue to bump steps past 4 or swap the seed before blaming the wiring.

Installing

Same pack as the Loader and Unload: search ComfyUI-Qwen-TeleStyle in ComfyUI Manager, or cd ComfyUI/custom_nodes && git clone https://github.com/Mister-Link/ComfyUI-Qwen-TeleStyle, then restart. The Loader article has the full first-run story - plan on a big one-time download before this node has anything to work with.