An idea, a minute of film,
and a sensible AI budget.

A skill, a model and a movie are different pieces

The starting point was this collection of video skills for coding agents. It covers a lot of ground: animated explainers, editing tools, avatars, motion graphics and generated scenes. A playable MP4 can come from any of those paths.

A skill gives an agent instructions and a process. A video model generates footage. An API host runs the model and bills for it. MCP gives the agent a standard way to discover and call tools. Connecting a model through MCP still leaves its compute and billing somewhere.

1. 01Idea & scriptWhat happens, and why?
2. 02StoryboardCharacters, shots, references
3. 03Generate & chooseSeveral takes, useful footage
4. 04Edit & soundPacing, continuity, final MP4

Remotion and HyperFrames are useful composition tools: they can turn authored code, images and clips into a video. Cinematic footage comes from a model or an existing source. A production system coordinates those pieces and keeps track of the decisions.

For a minute-long film, I would plan a sequence of short shots. Twelve five-second shots are an easy budgeting example. The real edit may need more footage for trimming, transitions and alternative coverage. A model’s short clip limit does not limit the length of the finished movie.

The shortlist that was worth inspecting

I looked for coverage from an idea to a finished file: script, shot plan, visuals, voice or sound, and assembly. Some projects cover that whole process; others supply one useful component.

OpenMontage was the broadest match for this particular goal. It has production plans, generation tools, editing and composition, along with a storyboard/review workflow. Its local and cloud backends make it possible to change the source of the footage without rebuilding the entire process. Its code is AGPL-3.0; the connected models and services have their own terms. Project documentation.

For an educational drawing-and-narration video, I would also consider explainroo. For a paper-collage explainer, Vox Director supplies a more specific visual language. The intended look matters as much as the list of integrations.

The two-3090 temptation

Assuming standard RTX 3090 cards, my hardware gives me two 24GB devices. The combined 48GB is useful, but a single-GPU application cannot automatically treat it as one 48GB card. The runtime must explicitly split the model or distribute the work. Two independent workers are another option; that is a throughput strategy, not a promise of exactly double the speed. CUDA device and memory behavior.

Wan 2.2 TI2V-5B is a documented starting point for 24GB hardware with CPU offloading. That establishes feasibility. It does not establish that the model meets my quality bar or that a one-minute film will be quick to produce.

At roughly nine minutes per shot, twelve shots would mean about 108 GPU-minutes for a first pass. That arithmetic is a planning scenario. Retakes, loading and editing add work; parallel workers and optimized runtimes can change elapsed time.

Local hosting alone does not make output worse, and cloud hosting alone does not improve it. The checkpoint and generation settings matter. The efficient 5B model is different from Wan’s larger A14B variants. I would establish the look I want with a few cloud samples before spending days tuning a local deployment. Wan model descriptions and reference implementation.

A useful local bridge: WanGP

WanGP is outside the original skills list, but it is relevant to a machine like mine. It supports several video, image and audio models, memory offloading, and a native MCP server. After installing it through its official guide, this starts its interface and MCP on a selected GPU:

python wgp.py --mcp-webui --gpu cuda:0

The default MCP endpoint is http://127.0.0.1:8000/mcp. Connecting a director skill to it still requires integration work. OpenMontage’s own Wan and ComfyUI adapters are the more direct starting paths. WanGP’s CLI and MCP documentation.

Other local paths include HunyuanVideo, CogVideoX and LTX. Larger LTX models require an appropriate quantized/offloaded workflow; two cards do not remove that requirement. “Open weights,” free software and unrestricted commercial use also mean different things, so I would check the chosen model’s license before building a production service around it.

OpenMontage leaves room to change services

A gateway such as fal or Atlas offers several models through one account. A direct provider API can offer different pricing or features. OpenMontage has both types of integration, plus local execution and self-hosted Modal endpoints.

This map describes implemented routes in the inspected source snapshot. It does not establish current entitlement on every account. A model’s name, version, resolution, reference support and audio settings must match the actual endpoint. ComfyUI Partner Nodes also call paid hosted services; putting one in a local graph does not make it offline.

What “state of the art” meant in this research

I used independent, blind-preference rankings to narrow the field. On October 15, 2026, MiniMax H3 at 768p led the audiovisual image-to-video boards at Artificial Analysis and Arena. On Artificial Analysis, Seedance 2.5 and Gemini Omni Flash 1.1 were close enough that their confidence intervals overlapped H3’s.

The same site’s silent image-to-video board placed H3 fourth, close to its leading group. That helps explain the choice beyond native sound. For text-to-video, Wan 3.0 led Artificial Analysis, while Arena’s ordering differed. The task and benchmark matter. Text-to-video board.

My conclusion is narrower than “H3 wins every scene”: H3 is a strong first candidate for high-quality footage at a relatively low API price. H3 Max and Turbo are distinct variants. A score for one cannot be carried over to another because their names are similar.

These rankings evaluate short clips, not a complete film’s continuity or story. For the final choice, I would compare a face/acting shot, an action shot and a camera move using the same references, then count how many takes I would keep.

Price the footage you have to generate

A $4.80 minute of generated output is not a $4.80 finished film. The number that matters is the cost of finding enough usable shots. I would track both the cash spent and the time spent waiting, reviewing and trying again.

“Three takes” means three candidates per required shot, including the accepted candidate: 180 generated seconds for a 60-second cut. It is a planning example, not a measured success rate. Separate images, billable references, agent usage, narration/music services, cloud rendering and taxes are additional. Native model audio is part of these models’ generated output.

Compare hosts as carefully as models

The Seedance row is fal’s published approximation, not Seedance’s global minimum. ByteDance’s official Ark sheet gives a no-video-input example of CNY 7.56 for five seconds at 720p: CNY 90.72 for 60 seconds before retakes. I keep that in yuan rather than conceal an exchange-rate assumption. Reference-video billing follows different token rules. Official Ark pricing.

Atlas advertises additional pricing routes, but a “from” price is not enough to establish an equivalent resolution, reference allowance or permanent rate. I would compare a quote for the exact request. The budget table uses verifiable endpoint prices, rather than an unsupported claim about the cheapest seller everywhere.

The easy-to-miss charges

MiniMax bills reference video as input as well as charging for output. Images beyond the included allowance can add charges too. A technically successful result that I discard still belongs in the creative budget; an API failure is a different case. MiniMax’s output and input pricing.

Self-hosting a model on a cloud GPU also has a meter. For example, ltx2-vidgen-skill’s advertised few-cents clip figure measures warm compute; its deployment keeps the GPU warm for 20 minutes. Loading and idle time are billable on Modal. The supplied 96GB GPU setup is also different from my local cards. Deployment settings and Modal billing.

Two adapter details change the cheapest route

The provider API and the integration were not quite in step. The official MiniMax API supports H3 at 768p and 2K, but the inspected OpenMontage direct adapter rejects anything other than 2K. Using the cheaper tier therefore needs an adapter update or a direct call whose result is imported. Pinned H3 adapter.

OpenMontage already has an H3 Max route. Its Turbo variant needs a separate route; the existing Wan-on-fal adapter similarly exposes Wan 3.0 rather than the inexpensive Wan 2.2 Turbo endpoint discussed during the research. The tool’s price estimate should also be checked against current provider billing. H3 Max route and Wan route.

For a first production, I would either enable H3’s official 768p tier or start with the existing H3 Max integration. I would test Turbo separately before trusting it for the shots that carry the film.

The first film I would make

I would begin with a small, bounded production rather than installing every tool in the catalog. The goal is a finished minute with a clear idea and consistent visual direction.

1. Write the story before spending on motion. Define the audience, tone, ending and spoken words. Turn that into a shot list with approximate durations.
2. Approve the references. Create character and location sheets, then a still frame for each important shot. Fix the look while changes are inexpensive.
3. Test the three difficult shots. Compare H3 and the cheaper Turbo candidate on acting, action and camera control. Record price, elapsed time and accepted takes.
4. Generate in short batches. Set a spending cap and review each batch before requesting more. Reuse references deliberately and leave enough footage for cuts.
5. Edit for a single film. Review faces, wardrobe, lighting, pacing and sound across shots. Assemble the cut, then handle titles, captions and the final mix.
6. Check the delivered file. Watch the export, listen through it, check caption readability and keep the prompts, takes and receipts with the project.

A useful brief for an agent would be:

Make a 60-second cinematic short from this idea. Propose the script, shot list and reference images first. Use image-to-video for the approved shots. Show the exact model, output tier and estimated cost before paid generation. Keep a log of every take and expense, stop at the spending cap, and deliver a reviewed MP4 with consistent sound and captions.

My two 3090s still have a place: reference images, supporting audio tasks, experiments and, eventually, a local workflow whose quality I already understand. I would decide where to run generation after seeing what earns a place in the edit.

What was checked

This is a dated research note, not a claim that I installed and benchmarked every model. The work checked project documentation and adapter code, independent rankings, official price sheets, a published GPU benchmark and publicly available sample media. It did not make paid generation calls or measure performance on my cards.

The OpenMontage code references are pinned to commit e45f5b12bd46. Prices and availability were checked on October 15, 2026. Before spending, I would recheck the chosen endpoint’s live quote and account access. The links throughout the note identify the sources behind the comparisons.