AIHub Expose Image Batch

AIHub Expose Image Batch is the heaviest node in the expose family, and the README says as much: "one of the most complex nodes meant to be used from video generation, handling, and lora training." Where AIHubExposeImage hands over a single layer, this one takes a series of images from the external app and packs them into a batched IMAGE tensor - the shape ComfyUI wants for video-to-video, image batch training runs, or generating several frames in one pass.

The mechanism is the same protocol as all the other exposes: an AIHub client (over the pack's websocket, port 8111) uploads the images, the server drops them into a temp directory and injects the file list as hidden inputs, and this node turns them into one tensor plus a MASKS tensor. Standalone in the web UI, it has nothing to grab. That's by design.

The inputs that matter

• type - all_frames, all_layers_at_image_size, or upload (default). The client uses it to know what to send: every frame of a video, every layer of the document at a uniform size, or a fresh upload batch.
• minlen / maxlen - the allowed batch size, so a client can't feed you 3 images when the workflow needs 16. You can even bind these to another expose's value via minlen_expose_id/maxlen_expose_id (plus offsets), same live-linking trick as AIHubExposeInteger.
• metadata_fields - the standout feature. This is a newline-separated mini-language that attaches per-image data to the batch. Each line is [field_id] [TYPE] [MODIFIERS...], e.g. frame_number INT SORTED MAX:total_frames or prompt_at_frame STRING MULTILINE MAXLEN:100. Types are INT, FLOAT, BOOLEAN, STRING; modifiers do sorting, uniqueness, and min/max validation (including referencing another exposed value like MAX:total_frames). The companion metadata_fields_label input supplies human labels for the same fields in order.
• normalizer - hand it the output of AIHub Utils New Normalizer to resize the batch to a consistent width/height before it becomes a tensor (video models like LTXV are picky about this).

Outputs

IMAGE (the batch), MASKS, WIDTH, HEIGHT, and METADATA - a custom AIHUB_METADATA object carrying each image's JSON payload. That metadata output is what makes training setups work: AIHub Utils Metadata Map can turn it into a single string (e.g. "0,48,97" for LTXV frame indexes), but the README is blunt that anything more complex - "like that on creating loras from image batches" - needs a custom node, because plain graph wiring won't cut it.

Install & gotchas

Same pack, same install: ComfyUI Manager → search ComfyUI-aihub-workflow-exposer, or git clone https://github.com/otavanopisto/ComfyUI-aihub-workflow-exposer into custom_nodes, then restart. No requirements.txt, no model downloads.

Where people get burned: the metadata syntax. A typo in metadata_fields - wrong modifier name, a MAX: pointing at an expose id that doesn't exist - and you'll chase phantom errors. Validate against the README's exact grammar, and remember metadata fields that reference another value (like MAX:total_frames) require that value to be an actually-exposed integer or project integer. Also: batches of wildly different image sizes will trip up downstream models; use the normalizer input rather than hoping the client sent uniform frames.