VLM Input Embedding Export

Most people never need to see the raw token embeddings their text encoder computes. If you're doing custom conditioning, probing attention, or building training data, you're the exception - and this is the node that hands them to you. UC_VLMInputEmbeds runs a Qwen3-VL encoder on your text and images, and exports the input embeddings as a 2D tensor and a state dict, or saves them straight to disk as a .safetensors file.

What it is

An embedding exporter for supported VLM encoders (the Qwen3-VL family via the pack's UC_Qwen3VLInputEmbeds base). You give it a clip, a text prompt, a list of image_paths, and a list of file_names, and it returns the raw embeddings the encoder would feed into its transformer - before any model layers run. That's the distinction to hold onto: these are input embeddings, not text-encoder output conditioning.

How it works

One important design detail up front: the prompt skips any template wrapping. The tooltip says it plainly - this is the raw prompt, tokenized directly, no <|im_start|> chat scaffolding. If you need the template, that's the job of a different node; this one is for people who want the bare tokens.

The image list and the file-name list must map 1-to-1, line by line. If you supply paths and names for the same images, each image gets tokenized at your chosen vlm_resolution (from the Fast 384 default up to 1536 and Original), and the embeddings are saved under models/embeddings using your file names (subfolders allowed, no .safetensors extension needed). slice_visual_tokens controls whether the first validated visual-token span is stripped from the output or kept in the full interleaved sequence - leave it off to preserve everything.

The node is safety-conscious in a way that matters: it validates every image path before starting any processing, so a typo aborts cleanly instead of half-writing files and leaving your VRAM in a weird state. And it calls clip.load_model() explicitly so the model registers properly as active.

The inputs and outputs that matter

• clip - a Qwen3-VL encoder.
• prompt - raw text, no template wrapping.
• image_paths / file_names - 1-to-1 line-separated lists.
• vlm_resolution - image analysis size.
• slice_visual_tokens - strip the first visual-token span (default off).
• Outputs: State Dict (a dict of {key_name: tensor_2d}) and Embeddings (the raw [num_tokens, hidden_size] 2D tensor).

Installing it

Part of ComfyUI-UtilsCollection:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart ComfyUI (or use Manager, search "ComfyUI-UtilsCollection"). No model downloads beyond the Qwen3-VL encoder you already load, minimal Python deps, and it needs a current ComfyUI.

Common issues

The count-mismatch error is the one everyone hits first - image_paths and file_names must have the same number of lines, exactly. The second is treating this as a conditioning node: it's not. It exports input embeddings for you to do something with; wire the outputs to a probe, a custom patcher, or just save-and-inspect. And mind the VRAM at high vlm_resolution - 1536 on an 8B Qwen3-VL alongside a diffusion model is a real ask. If you only need the saved files, don't bother keeping the outputs wired to anything heavy.