Qwen2 Run Inference

The other half of ComfyUI-ARZUMATA-Qwen2's vision pipeline. "Qwen2 Run Inference" takes the QWEN2MODEL object from the loader, an image, and a text prompt, and hands back the model's answer as a string - plus the original image, passed through untouched. That pass-through is the quietly useful part: you can drop this node into the middle of a graph, have it caption whatever flows past, and keep the pipeline moving without rerouting wires.

Why you'd bother

Batch captioning inside the graph, mostly. Feed it a stream of images (say, a dataset you're about to train a LoRA on), get natural-language descriptions back, and write them next to the images. Qwen2-VL writes sentences rather than comma-separated tags, which is exactly what the newer LLM-encoder generation models want for training captions - the same job JoyCaption and Florence 2 do, just with a Qwen model you already know how to load. If your target base is an anime tag model, this is the wrong tool; that's WD14 territory.

How it works

Under the hood the node builds a chat message - your image plus your text_input - converts the image to a base64 data-URI (using the same helper as the pack's Image2Base64 node), runs it through the Qwen chat template, and calls model.generate with the sampling settings you picked. Then it trims the prompt tokens off the output and decodes the rest. The min_pixels and max_pixels inputs control how the image gets resized before it goes in: they're pixel counts scaled by 28×28 internally, and the model clips the image to keep it inside that range while holding aspect ratio. Small range = faster, less VRAM, less detail. Bigger = sharper reading, costlier.

The inputs that matter

Required, you can't skip these:

• image - any IMAGE tensor.
• qwen2_model - from Qwen2ModelLoader.
• text_input - the prompt, default "Describe this image in great detail in one paragraph." It's multiline, so go ahead and write a whole instruction.

The optional knobs are the standard LLM sampling set, and the tooltips are the author's own words, so trust them: max_new_tokens (512 default, up to a million), temperature (0.7 - higher, more varied), top_p (0.9), min_p (0), top_k (0, meaning disabled), repetition_penalty (1). Defaults are sane. Leave them alone until output quality actually bugs you.

Two toggles do the real work for this node's actual use case:

• keep_model_loaded - defaults to false, meaning the model gets offloaded and VRAM cache emptied after every run. For a batch, flip it on, or you pay the reload cost per image.
• always_generate_captions - this is the one that saves you from ComfyUI's caching. ComfyUI skips re-running a node whose inputs haven't changed, which is great until you're captioning a changing stream of images and it decides nothing happened. Set this and it re-runs every time.

Outputs

• caption - the STRING the model produced. Wire it to a text-save node or anything that eats strings.
• image - the input image passed straight through, so your workflow doesn't dead-end here.

Gotchas

First inference on a fresh model is slow: the weights load into VRAM that run. If your captions stop updating, check always_generate_captions. If output reads robotic, nudge temperature up; if it repeats itself, repetition_penalty slightly above 1. And the usual Qwen-in-ComfyUI warning applies - the model is a several-gigabyte download that happens on first use, so a "hang" on the first run is usually just the download happening in the console.