AIHub Expose Project Latent

AIHub Expose Project Latent loads a stored latent tensor file out of the project and hands it to your workflow as a LATENT - no VAE encode round-trip, no pixels involved. This is the node for projects that keep intermediate results in latent space on purpose, and the pack's own example is video: for something like LTXV's step-by-step generation, storing latent files is far more reasonable than decoding frames to images and back between workflows.

The project setup is the usual one - a workflow has a project when its AIHubWorkflowController sets a project_type and the project is initialized; the client manages the files. This node asks the client for the latent file (default latent.safetensors), the server injects the resolved path, and the node loads it.

What you set

• file_name - the latent file as stored, extension included.
• batch_index - if the file belongs to a batch, pick one by index; a single integer, and negatives count from the end, Python-style.
• optional - true means a missing file yields a null instead of aborting.
• id - protocol field id.

Output is just LATENT, ready to feed a sampler or a VAE decoder.

The detail that shows this pack reads its own code

When the node loads the file, it does something subtle: it checks for a latent_format_version_0 key in the safetensors file, and if it's not there, it multiplies the samples by 1 / 0.18215. That 0.18215 constant is the SD1.5/SDXL VAE scaling factor. So a latent saved by older tooling gets unscaled back to the value range samplers expect, while latents carrying the format marker are left as-is. In practice: don't be alarmed if a latent you wrote out with plain SD1.5-era code comes back looking different from how you saved it - that's this node undoing the standard scaling. If latents look wrong downstream, the format marker (or its absence) is the thing to check.

Install & gotchas

Standard pack install: ComfyUI Manager → search ComfyUI-aihub-workflow-exposer, or:

cd ComfyUI/custom_nodes
git clone https://github.com/otavanopisto/ComfyUI-aihub-workflow-exposer

then restart. No requirements.txt, no model downloads.

Beyond the project-context gotcha (no project → no file), the real constraint is the same one that hits AIHubExposeLatent: latents are model-family-specific. A latent saved for one VAE won't silently work in a sampler expecting another's shape or scale. If you're chaining workflows through stored latents, keep them within the same model family - the node can fix scaling constants, but it can't fix "wrong latent for this model."