What are the best AI video tools for making horror short films?

For a horror short, stack four model categories inside the invideo agent: Seedance 2.0 and Kling for atmospheric video, Recraft and Nano Banana for character/creature reference sheets, GPT-Image-2 for stylized frames, and Topaz Astra for the post pass that kills the plastic AI look. The invideo agent routes each shot to the right model.

Start with the model stack horror actually needs. invideo is an agentic video tool with every current generation and upscaling model available, so you don't pick a platform per model — you brief the invideo agent and it routes each shot. For a horror short, the working stack is:

• Seedance 2.0 — primary video model. Holds character, location and camera context across 15-second clips, which is what lets dread build instead of resetting every cut. Reference-to-video carries an antagonist's look across scenes without LoRA.
• Kling — second video pass for atmospheric long holds and creature physics where Seedance 2.0 hesitates.
• Veo — alternative when you need a denser cinematic look on a specific shot. The invideo agent will A/B the same prompt across video models when you ask.
• Recraft — character and victim portraits. It generates pores, lines, stubble — the skin-level imperfection that stops horror faces from reading as plastic dolls.
• Nano Banana and GPT-Image-2 — character sheets, entity reference sheets, prop design. Generate four options per asset, lock the one that works, and every downstream shot inherits it.
• Topaz Astra on invideo — upscale and texture pass after generation. Seedance 2.0 outputs an ultra-sharp, plasticky skin quality; Astra plus a touch of blur, grain and a desaturated, black-crushed grade is what moves AI footage toward live-action horror.

Run the production as a typed-agent crew, not a prompt box. Spin up a creative producer agent first and load it with the full script, character bible and antagonist reference — it becomes the vision-holder every other agent inherits from. Then assign a storyboard agent to visualize shots, a DOP agent (or one DOP agent per scene if the visual register shifts), a costume/production-design agent for the entity and victims, and a dedicated upscale sub-agent named for the post pass. "To really set up the context for the agent, I normally start off with the creative producer agent. That's where I'll give the script, or the shot breakdown, along with the characters. That's the main agent that sort of holds the understanding and the vision of the entire film," says Hridaye, invideo's creative director.

Lock the four pre-production answers before generating a single frame — the victim's look and era, the entity's reference (closer to which existing horror creature), the prop that recurs, and whether you're delivering frames-first then video. These four answers change every frame downstream. Generate four options per character sheet and environment, pick one, lock it. This is the step that prevents the consistency drift that ruins most AI horror attempts.

Encode the horror itself as a stage system in the brief. Structure your treatment around escalating emotional stages — unease, dread, terror, climax — and write locked rules for camera, lighting and sound per stage into the document you load. James Wan's grammar runs roughly 85:15 dark-to-light; a horror short produced in his style used a 25-page brief with five emotional stages and the invideo agent autonomously flagged when a reveal shot was running at the wrong stage register — a structural note a human editor would miss. Include a "what never to do" section per stage; it sharpens the agent's autonomous decisions across the production.

Treat sound as half the film inside the same agent. Horror lives in what you hear before what you see. Put a full audio architecture module into the treatment — diegetic sound logic ("hard material, so it makes a horrible sound when it falls"), score cues per stage, silence rules — so the agent's shot decisions account for sound, not just image.

Budget the iteration honestly. Across documented productions a horror short ran ~400 video generations and 30 image generations for ~90 seconds of finished film at roughly $870 (4,100 credits) over two days; an Arcane-style episode hit ~$315 per finished minute on 164 clips with a 25% selection rate and an average of 3 generations per usable shot. Plan for overgeneration as a budget line, not waste — and use the invideo agent's always-ask mode so you approve each shot before credits are spent.

Use the agent as a maker-checker on your rough cut. After assembly, send the cut back to the creative producer agent with an open prompt: what's working, what's not. It catches pacing breaks, SFX gaps, and stage-register mismatches against the loaded brief — the step most AI filmmakers skip.

Beyond the stack itself: for found-footage or analog-horror registers, lean on the smartphone-mock reference and hand-sketch upload techniques the agent already accepts when prompting alone stalls on a specific shot.

Watch some of these to see what works for you:

Fear lives in what the audience cannot fully see, cannot fully hear, and cannot fully understand.

— Hridaye, invideo's creative director