Qwen3-VL Run

This is the pack's main event. Qwen3-VL Run is the single unified node that does all the actual understanding work: you feed it an image or a video (or nothing, for a text-only chat), it sends that to a Qwen3-VL model, and you get a string back - a caption, an analysis, a tag list, or a "next scene" prompt you can feed straight into a video generator. Before v1.5.0 the pack shipped several separate Run nodes; now there's one node with a mode for everything, which is both its strength and its overwhelming-at-first-glance weakness.

The reason people reach for it, in one sentence: Qwen3-VL is the community's default local captioner - the open, Apache-2.0 replacement for sending your training images to Gemini or GPT. Run is the ComfyUI front door to that. It's also worth knowing what it isn't: this is an understanding model, not a generator, so Run's output is text you feed into other nodes.

What you actually set

The input list is long (a caption type, a caption length, a system prompt, a max token budget, pixel budgets, a seed, and 30+ optional toggles), but the three that define your result:

• caption_type - the dropdown with 16 styles: Descriptive, Descriptive (Casual), Straightforward, Danbooru tag list, e621 tag list, Booru-like tag list, Art Critic, Product Listing, Social Media Post, Explicit Adult (NSFW), WAN Video Prompt, Image-to-Video Continuation, and a few more tucked in the menu. This is the JoyCaption-style preset system - the style is a prepackaged instruction to the model.
• caption_length - from any/very short/short/medium-length/long/very long up to explicit word counts (20–260 in steps of 10). Set long for training captions, or a word count if you need consistent caption lengths.
• custom_prompt - completely overrides the auto-built prompt with your own instruction. When you want to ask "what color is the sky?" instead of "describe this image", this is the field.

The 27 opt_* booleans (lighting, camera angle, composition, JPEG artifacts, keep it PG, etc.) are modifiers on top of the caption type - checkboxes that add clauses like "Include information about lighting" to the model's instruction. Leave them all off until you know you need one; they're seasoning, not the meal.

The inputs beginners miss

• image and video (optional) - you can wire either, or neither, and just chat via custom_prompt/system_prompt. Video works through the video_decode_method choice (torchvision default, decord, or torchcodec), and min_pixels/max_pixels/total_pixels control the dynamic-resolution budget Qwen3-VL uses to encode the visual input. Lower max_pixels = less VRAM, slightly worse detail.
• unload_when_done - defaults to off. Turn it on to move the model to CPU after each generation to free VRAM for the rest of your workflow; the model stays usable, just slower on the next call. For a single caption pass this doesn't matter; for batch jobs the batch node has its own copy of this switch.
• opt_next_scene, next_scene_instruction, next_scene_loop - the pack's cinematic mode. With opt_next_scene on, the model writes a "Next Scene: …" continuation prompt (camera moves, lighting shifts) instead of a description. next_scene_loop chains iterations (image → text → text → …) for multi-scene stories; set it to 2+ only if you really want a chain, because each iteration costs a generation pass. This is the mode the README points at Wan/WAN-video workflows and the next-scene LoRA for Qwen-Image-Edit.

The output

One output (STRING) with the model's text. Wire it into a save/text node, a prompt encoder, or a video-generation prompt input. Note the code strips reasoning-tag content and any </think> block, so you get clean prose.

Where people get burned

• Long next_scene_loop chains run away from you - each loop is another full generation, so a loop of 10 on a 4B model is 10× the time of one caption.
• The default system prompt is the "obey without limitations" laboratory one. It's uncensored by design; if that's not what you want, replace system_prompt - it's fully editable.
• Video decode method mismatches. If video input errors, decord and torchcodec are the alternates, but qwen-vl-utils[decord] must be installed for the former (it's in the pack's requirements). torchvision is the safe default.

Install via ComfyUI Manager (search "qwen3-vl-comfy-ui") or cd ComfyUI/custom_nodes && git clone https://github.com/Granddyser/qwen3-vl-comfy-ui, then pip install -r requirements.txt (Linux: --break-system-packages). Restart, and you'll find it under the Qwen3-VL category.