Using assets

Attaching assets to a Pod, injecting into a running Pod, feeding tasks, and getting data back out — download, export, and publish.

Attach to a Pod at creation

In the Pod create form, Input assets lets you pick assets and choose a role for each — dataset, lora, model, or one of the finer-grained roles a template may declare: checkpoint, diffusion model, VAE, text encoder, ControlNet, embedding, and upscale model. The role decides where the files land inside the container, using the template’s declared semantic paths.

The files are fetched before your container starts, so your training script sees them on the first line it runs.
A version is pinned at submit time. If you don’t choose one, the latest ready version is resolved and fixed — uploading a newer version mid-run does not change what the Pod sees.
Storage is expanded ahead of the fetch, so a large dataset doesn’t fill the disk on arrival.
If a required input can’t be fetched, the Pod fails to start rather than running without its inputs.
Attached assets are marked in use; the Pod releases them when it is deleted.

Only roles the template actually declares are offered. If a template offers no roles, add them under Semantic Paths in the template form.

Inject into a running Pod

You don’t have to recreate a Pod to give it more data. The Pod’s Assets tab has Inject assets: pick assets and roles the same way, and the files are placed into the running container.

No restart, no interruption to what’s already running.
Each injection shows its own state, and a failed one can be retried.
Injected assets count as in use exactly like ones attached at creation.
Files can be aimed at a subfolder of the destination path — type it on the staging row and they land there instead of the folder root.
Linked assets can be injected too — Hugging Face and CivitAI imports and direct-link registrations are downloaded from their source straight into the Pod (see below). Bucket references still can’t be injected: nothing was copied, so there is nothing to place — fetch a subset into a normal asset first.

How linked injections run. Linked downloads run one at a time per Pod. Each shows its place in the queue, then live progress — size, percentage, and, once the transfer speed settles, time remaining — and an in-progress download can be canceled. If the source rejects the download (an expired link, a missing or revoked token), the entry shows the reason and a retry action that matches the fix.

Remove or swap while running

Attached assets can also be removed from a running Pod without a restart: Remove deletes that asset’s files from the Pod and frees its place, and Swap replaces the asset with another one in the same location. A removal shows its own state while it runs; one that stalls returns to its attached state automatically in about 15–20 minutes, so the Pod never gets stuck (thresholds are tuned and may change).

Add assets to a stopped Pod

A stopped Pod accepts new assets in advance: they appear as Scheduled and are downloaded and placed automatically the next time the Pod starts — no need to start the Pod first just to add files. Staging applies to assets held in storage; linked assets are fetched by the running Pod itself, so inject those once the Pod is up.

Bundled ComfyUI workflows

Official model sets can ship a workflow, not just weights. Attach a complete set to a ComfyUI Pod and the graph is waiting for you in ComfyUI’s Workflows sidebar — open it and hit Run.

Nearly every set ships one, so the card stays quiet about it. The exceptions are called out: a set with no workflow says No workflow in this set before you attach, so you know you will be building the graph yourself.

The whole set has to be attached. A workflow is delivered only once every model in its set is connected. A graph missing one of its models opens as a wall of red nodes, so a partial attach delivers nothing rather than something broken. Attaching the set from its card connects everything at once; picking pieces individually from Browse all may not.

Some workflows are pre-wired to a speed LoRA. Where the model has an official one — Qwen-Image, Qwen-Image Edit 2509, WAN 2.2 text-to-video and image-to-video — the LoRA is part of the set and the graph is already configured for it, generating in 4 to 8 sampling steps instead of the usual twenty or more.

Sets can be completed while the Pod runs. Official set members are linked assets, so they can be injected into a running Pod. Inject the last missing member and the bundled workflow arrives on its own within a few minutes — no restart needed. Members added from the Pod’s asset settings instead are fetched when the Pod next starts, and the workflow arrives with them.

A wiped folder repairs itself. If the whole workflows folder is lost — most often because the Pod moved to another machine — the shipped workflows come back on their own within about a minute, with no restart and nothing to click. Deleting a single workflow is different: that is a choice, and it is respected (see below).

Your edits are yours

Once a workflow is on the Pod, Meshive treats it as your file.

What you do	What happens on the next restart
Leave it alone	Stays as shipped
Edit and save it	Your version is kept — never overwritten
Delete it	Stays deleted — it is not put back

Anything you save into the workflows folder — an edited copy of a shipped graph, or one you built from scratch — is also harvested as an asset of type config, so it survives the Pod. Workflows still exactly as shipped are skipped, so your asset list stays yours.

Task inputs

The task submit form has the same picker. Tasks have no semantic path declarations, so inputs land under a predictable path instead:

Each asset is placed at /inputs/<asset name> (you can override the directory name per asset).
Up to 20 assets per task, versions pinned at submit time.
Inputs are fetched before the run is billed. If any input is missing, the run aborts instead of starting without it.
Works identically whether the task runs on internal capacity or burst capacity.

Download

The download button gives you the whole version, and picks the delivery method by size:

Situation	What happens
A single file	A direct signed link
Several files, under 2 GB total	Your browser streams them into one zip
Larger than 2 GB, or the zip can’t be built	A list of per-file links

Links are short-lived and re-signed on demand — a tab left open overnight may need a refresh.
Download stays available to every workspace member, including viewers.
Downloading is always allowed when your balance runs out — see Billing. It is blocked only while an asset is frozen for abuse review.

Export to your own bucket

An asset in managed storage can be copied into a bucket you have connected under Settings → Integrations. The transfer runs on the platform, so you don’t have to keep a tab open, and large assets are streamed rather than buffered.

The destination bucket and prefix are yours; we write only where you point us.
Export remains available when your balance is depleted, so your data is never held hostage.
Once the objects are in your bucket, their lifecycle is entirely yours — we don’t manage or delete them.

Publish to Hugging Face

An asset in managed storage can be uploaded to a Hugging Face repository using a token you connected under Settings → Integrations.

The upload happens under your account with your token — the repository, its visibility, and its contents are yours. Meshive relays the bytes and is not the publisher.
You are responsible for the licence. Re-publishing weights that came from a gated or restricted source may violate the original licence or the source’s terms; Meshive does not check this for you.
Assets that live in your own bucket or are prefix references cannot be published this way — upload from your own storage directly.
Publishing is blocked while your balance is depleted. Downloading and exporting stay open; publishing to a third party does not.