Qwen3-VL Loop Run (Advanced)

Qwen3-VL Loop Run (Advanced) is the same vision-language engine as the plain Run node, but rebuilt for being driven by a loop rather than sitting at the end of one. It captions or analyzes a single image per call, and - the part that makes it a loop citizen - it returns next_index and is_last so whatever is looping over your workflow knows whether to continue and with what counter. If Batch Run is "process everything, then I'll tell you," this is "process this one, tell me what's next."

It's the "Advanced" label, and it earns it in one specific way: this node does not know about folders at all. You give it one image per invocation, so you pair it with Qwen3-VL Folder Loader (Loop) (which supplies the image plus total_count) or with any other per-iteration image source. The loop controller - ComfyUI's built-in loop nodes or whatever you use - drives loop_index and total_loops, and the whole thing walks your folder one caption at a time, with the option to pipe each prompt into downstream generation.

The mode switch is the headline

• mode - a 6-way dropdown that changes the prompt entirely: next_scene_creative (default - the model invents a cinematic continuation), next_scene_guided (continuation shaped by your custom_instruction), video_continuation (image-to-video style), wan_video (WAN-format video prompt), descriptive (plain captioning), and custom (your instruction verbatim). This is the pack's "Next Scene" system in loop form - the mode that writes camera moves and scene evolution prompts for feeding into Wan/Luma video generation. If you just want captions, set descriptive and ignore the cinematic stuff.

The rest of the required list is familiar if you've used Run: system_prompt, max_new_tokens (512 default, up to 4096), min_pixels/max_pixels, seed, and the same 27 opt_* toggle checkboxes. custom_instruction matters when mode is next_scene_guided or custom; leave it empty in the creative mode and the model runs free.

The four outputs - the loop contract

• prompt (STRING) - the generated text for this image. What you actually want.
• next_index (INT) - just loop_index + 1. Feed this back into your loop controller (or into the Folder Loader's index) to advance.
• is_last (BOOLEAN) - true on the final iteration (loop_index >= total_loops - 1). Stop the loop on this.
• status (STRING) - a human-readable Loop 3/20 string for display/logging.

How it behaves in a run

The model is moved to GPU when the loop starts and - with unload_when_done defaulting to true - moved back to CPU only after the final iteration, so mid-loop iterations don't pay a reload tax. That's the whole point of the loop flavor: keep the model resident, iterate, unload once. The console prints progress (Loop 3/20 (FINAL)) per call.

Where people get burned

• Wrong mode = wrong output. The default is next_scene_creative, so a first-time user who just wants captions gets cinematic "Next Scene:" prose and wonders what broke. Set descriptive (or custom with your own prompt) for captioning.
• Your loop controller has to speak index/total_loops. This node doesn't self-advance; if your controller doesn't feed loop_index and read next_index/is_last, the loop won't move. Wire the outputs before blaming the model.
• Per-iteration VRAM on tiny cards. It keeps the model resident through the loop, which is faster but means the model stays in VRAM the whole time - if the rest of your workflow needs that VRAM mid-loop, unload it yourself between sections.

Install via ComfyUI Manager (search "qwen3-vl-comfy-ui") or clone https://github.com/Granddyser/qwen3-vl-comfy-ui into ComfyUI/custom_nodes + pip install -r requirements.txt, restart, and find it under Qwen3-VL. Loop Run is the advanced choice - Batch Run is simpler if you don't need per-image downstream work, and the two share every prompt-setting input so switching between them is painless.