TTP QwenVL3 Local Loader

A small, single-purpose loader node that exists because of one specific limitation in TTP Smart Tile Interactive Crop: that node's Auto Tile feature can use QwenVL3 to find objects worth tiling (faces, hands, text), but it doesn't read a QwenVL .safetensors file directly. This node is the missing link - it loads the model separately, and you pass its output into Interactive Crop's qwen_vl_model input.

What it does. model_file is a dropdown populated from whatever's sitting in ComfyUI/models/text_encoders - the pack's own default placeholder text, put_qwenvl3_model_in_models_text_encoders.safetensors, is a literal instruction: if that's all you see in the dropdown, it means the folder is empty, not that something's broken. Drop a QwenVL3 checkpoint in there (the README's own worked example uses qwen3vl_4b_fp8_scaled.safetensors) and refresh the node to see it listed. model_family (default auto, alternative qwen_vl) and device (default vs. cpu) round out the inputs - device=cpu is there for when you'd rather keep the tagging pass off your GPU, which is a reasonable trade since it only runs during the Auto Tile detection step, not during actual sampling.

Outputs: qwen_vl_model (the TTP_QWENVL3_MODEL object Interactive Crop wants) and info (a status string).

Why this matters for Smart Tile. Per the pack's README, QwenVL3-based Auto Tile is one of two detection paths (the other is SAM3.1) for building an object-aware tile layout automatically instead of drawing every tile by hand. QwenVL returns bounding-box JSON describing where it found things - Smart Tile accepts several common formats (bbox, bbox_2d, box_2d, objects, xywh) - and if the model returns one big full-frame box instead of individual objects, Smart Tile is smart enough to split that into useful detail tiles rather than treating the whole image as one tile. Results get cached by model file, tile hash, prompt, and seed, so re-running a workflow doesn't re-interrogate tiles that haven't changed.

Install. ComfyUI Manager, search "Comfyui_TTP_Toolset", or cd ComfyUI/custom_nodes && git clone https://github.com/TTPlanetPig/Comfyui_TTP_Toolset, then restart. The real dependency here isn't the node code, it's the model file itself: you need to separately download a QwenVL3 vision-language checkpoint and place it in ComfyUI/models/text_encoders - this node doesn't fetch or ship one for you. Expect this to be a multi-gigabyte download depending on the variant you choose (the fp8-scaled 4B variant referenced in the README is the lighter end of that range).

Common issues. If the model_file dropdown only shows the placeholder text and nothing else, that's the node correctly reporting an empty folder - put a .safetensors file in models/text_encoders and restart or refresh the node list, don't go looking for a separate error. If Auto Tile in Interactive Crop still isn't detecting anything after you've loaded a model here, double-check the wiring: this node's qwen_vl_model output has to actually be connected to Interactive Crop's qwen_vl_model optional input, and Interactive Crop's own auto_detect_mode widget has to be set to qwenvl3 - both are separate steps, and missing either one leaves Auto Tile silently doing nothing.