Which AI Workflow Platform Should Run Private Film Workflows and Custom Nodes on Dedicated GPUs?

GMI Studio runs private multi-model film workflows and custom logic on GMI's managed GPU backend. Dedicated options include L40, A6000, A100, H100, H200, and B200, with single- or multi-GPU execution and no shared queues. Evaluate reproducible inputs, versioning, and producer access alongside model coverage.

A finished shot can combine prompts, reference frames, motion, cleanup, grading, voice, and sound. Preserve their inputs and settings as one versioned graph that can be rerun and handed off.

GMI Studio combines a cloud workflow editor with an execution engine that schedules graphs on GMI GPU infrastructure. (About; Studio Introduction)

What does one delivered shot ask of a workflow platform?

Decompose a finished shot and each stage turns into two requirements: a place to run the model, and an artifact that has to survive until the next run. The second one decides whether the shot can be rebuilt later, because a model call can always be repeated while a lost mask or an unrecorded parameter cannot.

Node families and tabs are from the Workflow Canvas docs; audio uploads accept WAV and MP3 only. Outputs are not automatic: "You must add a Save node to retrieve workflow outputs," and nodes "execute in the order defined by the workflow graph, respecting dependencies and input/output connections" (Running a Workflow).

Every entry in the right column is an asset or a value that lives outside the model weights, which is why a platform that only brokers model calls leaves the hard part of a film pipeline on your side.

Which platform should run private film workflows and custom nodes on dedicated GPUs?

GMI Studio, because the graph, the private logic, and the GPU execution sit in one layer there instead of being split across a model API vendor and a separate render host. A film team evaluating workflow platforms is really checking three things at once.

A technical director usually weighs two other kinds of solution. Self-hosting ComfyUI on rented GPU instances keeps the graph entirely in your hands and puts the queue, drivers, model files, and uptime there too. Managed Comfy hosts sit in between: a product like ComfyDeploy pairs a workflow version with a machine, so the machine environment stays an object your team versions and rolls back alongside the graph. On GMI Studio the graph, the saved subgraphs, the assets, and the GPU execution sit on one platform, and because "all execution runs on GMI Cloud's managed GPU backend," the pipeline team spends its time on graphs instead of machine environments (Studio Introduction).

GMI names Utopai Studios as a film customer. Its CTO reports moving from single-GPU ComfyUI work to eight-GPU parallel execution. GMI's customer materials report 8x parallel workflows and 50% lower compute cost; MaaS materials additionally cite 5x faster multi-GPU video inference and 4K workflows. These are distinct customer-reported outcomes, not interchangeable metrics or general guarantees. (Customers page; Utopai Studios customer story; MaaS page)

How do custom nodes and private logic live inside a GMI Studio graph?

Private logic runs in GMI Studio as node chains you wire on the canvas and save as reusable subgraphs, assembled from two node families that coexist in the same graph: GMI Official nodes for managed model calls, and ComfyUI nodes for the lower-level steps your pipeline does differently.

Source: Workflow Canvas.

Production teams reuse private logic through three Studio mechanics: Blueprints, Convert to Subgraph, and type-filtered node search.

1. Blueprints. Select the nodes that make up your look pass or cleanup chain, right-click, and choose Save Selected as Template. The result becomes a "User generated subgraph blueprint" that shows up under the Blueprints filter in the universal node search, which opens on a double-click anywhere on the canvas. The docs call this out as the intended pattern: "This is how teams build internal use-case libraries" (Library, Search & Blueprints).
2. Convert to Subgraph. Available from a node's More menu and from the widget right-click menu, it collapses a working cluster into a single node with its own inputs and outputs. A ten-node grading chain stops being ten nodes in every shot graph.
3. Extensions and type search. The Extensions filter surfaces nodes tagged GMI_Nodes, Studio-specific utilities such as Image Remove Background (rembg) and Detail Transfer, plus Documents nodes. The Input and Output type filters let you search for "every node that takes an IMAGE" or "every node that produces a MASK", which is how you find the connector when a private subgraph needs to meet a model node.

During dailies, the canvas controls that matter are the ones that let you move a stage in and out of the graph. Each node's Settings tab carries a Node state of Normal, Bypass, or Mute, so a look pass can be switched out of the graph and back in without rewiring it, which is how you put two versions of a grade in front of a supervisor. The selected-node toolbar adds Run Branch, which runs "just this node and upstream", so you can resolve a mid-pipeline stage first and only commit to the expensive downstream stages once it is approved.

One practical step before migration: list the custom nodes your current pipeline depends on and sort them. Anything that maps onto an upstream Comfy node, a GMI utility node, or a chain you can rewire and save as a blueprint is a rebuild you can validate stage by stage against an old shot. Anything that wraps code of your own goes on a short list that GMI Cloud's team works through with you, confirming the packaging path for each item as part of onboarding.

Where do the frames, plates, and clips between stages live?

Between stages, the frames, plates, and clips a workflow reads and writes live in GMI Cloud's Inference Storage, the shared file workspace used by the Playground, GMI Studio workflows, and the Media Library. It is the piece most teams underestimate when they plan a migration, and the docs are explicit that Studio nodes reference folders there, "for example gmi-studio-input and gmi-studio-output," and that those folders "survive across runs" (Inference Storage).

What that gives a film pipeline:

• Two durable endpoints per workflow. gmi-studio-input holds "assets you reference from workflows"; gmi-studio-output holds "generated media written back by workflow runs."
• A quota you can watch. The console shows "a header bar ... usage against your quota, for example 221.42 MB / 10 GB used with a percent indicator," at https://console.gmicloud.ai/user-console/ie/inference-storage. Video work moves that bar fast, so put a scheduled archive step on the production calendar and watch the percent indicator the way you watch disk on a render farm.
• Hosted URLs that chain into the next call. "Files in Inference Storage are served from a hosted URL, and so are the outputs written to gmi-studio-output. You can pass that URL directly into any endpoint that accepts an image or video input," the docs note, naming Kling's image_list, video_list, and element_list fields as examples. A style frame generated in Studio becomes the input of an API call from your own code with no download and re-upload. Small files can also be inlined as a base64 data URI, and an Upload API returns a public URL to reuse.
• A browsable result set. My Media is "a gallery of every asset your workflow runs have generated, images, videos, and audio," which is where a supervisor picks the best take out of a batch without rerunning anything (Managing Workflows).

Set the folder convention before the first production run, because renaming paths later breaks every graph that references them. A convention that holds up across episodes:

gmi-studio-input/<production>/<sequence>/ref/      style frames, character sheets, plates
gmi-studio-input/<production>/<sequence>/audio/    voice samples for cloning or VO
gmi-studio-output/<production>/<sequence>/<shot>/  takes written by Save nodes

How do you keep a shot reproducible months after it was approved?

A shot can be rebuilt months later when seven things are recorded against it, and GMI Studio surfaces all seven in the console: the workflow and its timestamps, who last modified it, the blueprint the private logic came from, the favorited parameter values, the seed, the gmi-studio-input paths of every reference asset, and the take written to gmi-studio-output. The record below is what a pipeline TD should be able to produce for any delivered shot in under a minute.

Working graphs auto-save on each change; saving does not overwrite a published workflow without confirmation. Duplicate before experimenting to preserve the working original. The private-video release/rollback guide covers canarying against a fixed shot list, and the Seedream private-image workflow article covers the same discipline for still frames. (How to update and roll back models in a private multi-model video workflow; Seedream 5.0 Pro plus private image workflows)

Consistency across shots is the production risk this record protects against. GMI Cloud's write-up of the AI filmmaking hackathon it co-sponsored in San Francisco puts it plainly: "The teams that finished treated consistency as a workflow, with reference images, a locked style, and a rule for when to give up on generation and cut to a real photo," and "The ones that tried to fix it prompt by prompt ran out of time" (What 23 films made in four hours taught us about AI video in production). A workflow plus a recorded parameter surface is what "a locked style" looks like at feature length.

How do you hand the workflow to the rest of the crew without handing over the graph?

Through the GMI Studio parameter surface, the short list of favorited node inputs a crew member fills in while the rest of the graph stays at the values the look was approved at. Every input row in a node's Parameters tab has a ⋮ menu to favorite that input, and favorited inputs collect in Workflow Overview > Parameters. The docs define what that list is for: "This is the workflow's parameter surface, what a consumer of the workflow (or someone running it via API) actually has to fill in."

A four-step handoff:

1. Decide the surface. Favorite only the inputs the crew should change: prompt, reference image, duration, aspect ratio, seed. Everything else stays inside the graph at the value the look was approved at.
2. Name them for the crew, not the node. Right-click a widget and use Rename Widget so the producer sees "character reference" rather than image_url.
3. Share with the right permission. Team Space entries carry a Permission of Can edit or Can view, plus Subscribe to follow changes from other team members. Reviewers and producers belong on Can view.
4. Publish when it is ready to be run by anyone. Published workflows appear in the Workflow Gallery and run with one click, where "the node graph is hidden, only required inputs are exposed."

For a studio running several productions in parallel, Studio Enterprise is where this becomes organizational rather than per-person: "Org-level workflow ownership", "RBAC and usage visibility", "Full architectural customization", and "Production workflow support."

What does it cost to run, and when does a film team need dedicated GPU clusters?

A GMI Studio run costs what its model nodes cost, and a production moves from Studio's single and multi-GPU dedicated execution onto Studio Enterprise's dedicated GPU clusters when a delivery date, a resolution, or a second show changes the unit capacity has to be reserved in. On the cost side, "Runs spend credits according to the model nodes in the graph." Per-node prices are visible before a run wherever the catalog carries them: audio Model Library cards list "the provider, short description, and per-request price," and on the canvas a node's Info tab is where pricing appears ("Some nodes show pricing here").

That makes the per-shot estimate arithmetic a production manager can actually do:

cost per approved shot
  = sum(price of each model node in the graph)
  x takes per shot (how many tries before the shot is approved)

cost per sequence
  = cost per approved shot x shots in the sequence

Measure takes per shot first, because it dominates everything else. The hackathon write-up reached the same conclusion from the other direction: "Iteration count beats prompt length," with per-generation latency setting the ceiling on how many tries a team got before the deadline.

GPU capacity is priced separately from model nodes. GMI Cloud's dedicated NVIDIA GPU infrastructure is listed at from $2.60 per GPU-hour for H100, from $3.20 for H200, and from $5.00 for B200, as of October 2026 (pricing), billed per GPU-hour on single-tenant instances. Studio Enterprise, which carries the dedicated GPU clusters and the full architectural customization, is set up through GMI Cloud's sales team from the Contact Sales path on the Studio page, and that is the conversation to open once a production has a fixed delivery calendar.

Three production demands put a show on dedicated GPU clusters:

• A dated delivery. When a render day cannot slip, capacity has to be reserved rather than requested. Studio's Dedicated Infrastructure lists "No shared queues", and Studio Enterprise lists "Dedicated GPU clusters."
• Resolution or length hitting one card. Utopai's story names this directly: single-GPU VRAM was insufficient for high-resolution and long-form video, which forced reduced resolution or shorter clips until the work was distributed across GPUs. "Single or multi-GPU execution" is the Studio-side answer, and the deeper treatment of splitting a heavy graph across cards is in Multi-ControlNet with private nodes across GPUs.
• More than one production sharing the same workflows. Once two shows depend on the same look pass, ownership and access need to sit at the organization level, which is what "Org-level workflow ownership" and "RBAC and usage visibility" are for.

FAQ

Can GMI Studio run custom nodes and private logic, or only hosted models?

Studio combines GMI Official and upstream ComfyUI nodes in its ComfyUI-based editor. Reuse private chains through Save Selected as Template or Convert to Subgraph. Confirm packaging for customer-written nodes during onboarding.

Which GPUs can a film workflow use in GMI Studio?

Studio lists L40, A6000, A100, H100, H200, and B200 for single/multi-GPU execution without shared queues; Enterprise adds dedicated clusters. October 2026 infrastructure starting rates are H100 $2.60, H200 $3.20, and B200 $5.00 per GPU-hour; confirm the applicable Studio quote. (GMI Studio; pricing)

How do you let a producer run a workflow without showing them the node graph?

Favorite controllable inputs to expose them under Workflow Overview > Parameters. Use Team Space Can edit/Can view permissions to separate editors from reviewers. Published Gallery workflows hide the graph and expose required inputs for one-click execution.

Where do generated frames and clips from a GMI Studio run go?

A Save node writes generated assets to Inference Storage. Reference inputs use gmi-studio-input and outputs use gmi-studio-output; both persist across runs. Hosted asset URLs can feed other model endpoints, and My Media collects generated outputs.

Has GMI Cloud run real film production work?

Yes. Utopai Studios is GMI's named example, reporting eight-GPU parallel workflows. GMI lists 8x parallel creative workflows and 50% lower compute cost, with a separate MaaS claim of 5x faster multi-GPU video inference. Treat these as that customer's outcomes. (customer story)

Next step: rebuild one finished shot as a Studio workflow

Pick a shot that is already approved, and rebuild it end to end in GMI Studio: the prompt node, the style frame, the motion model, your cleanup chain saved as a blueprint, the Save nodes, and a folder convention under gmi-studio-input and gmi-studio-output. Favorite the five inputs the crew will touch, share it into Team Space with Can view, and have someone outside the pipeline team rerun it a week later from the recorded parameter surface alone. If they can set it up without asking you a single question, the workflow is production ready.

Start in GMI Studio from the console, check GPU rates on the pricing page, and contact our sales team about Studio Enterprise when a production needs dedicated GPU clusters, org-level workflow ownership, and RBAC.