What top AI filmmakers actually use

Counts reflect how many of the eleven contributors named each tool across their stack, their strong views, and their lean setup. Nobody uses all of these at once. The recurring theme was that sustainable AI filmmaking is a small repeatable pipeline, not a drawer full of subscriptions.

Script and story

Claude is the brain of the operation for most of the panel. The pattern is consistent: start with a free flowing conversation about the concept, ask the model to push back, then work it into a structured script. Several said plainly that they reach for Claude over ChatGPT when the work gets complex, citing longer context and a better feel for tone. One filmmaker runs research and storytelling through Perplexity first, then switches to Claude for the writing itself.

Images

Nano Banana and Nano Banana Pro are the default, named by ten of eleven, generating the scene and character assets once the look is locked. But this stage is more contested than the count suggests. A camp inside the panel prefers Higgsfield Cinema Studio for first frames, because it lets you specify focal length, lens, aperture and even camera hardware in plain language. One contributor switched to GPT Image 2 for first-frame generation for a sharper reason: with a recurring lead character, other tools drift on likeness across scenes, and he needed the same face in every still, not a stylistic cousin of it. And Midjourney survives with five of the panel for one specific job: a distinctive visual style that the other generators tend to flatten.

The hub: a genuine split

Two platforms compete for home base, and the panel divides almost exactly in half. The Magnific camp loves its node-based Spaces environment, described as feeling like a visual canvas rather than a prompt box, with the newest models available the day they land. The Higgsfield camp loves it as an access layer, one place to run the same prompt through different engines, with new creator-facing features shipping constantly. The disagreement gets specific on cost: one contributor calls Magnific's credit pricing reasonable, another says its upscaling has become expensive and occasionally unusable, sending her back to Topaz. Both are right about their own workloads, which is the point: the hub choice depends on whether you want a building environment or a switchboard.

Sound and the cut

ElevenLabs is the only voice tool the panel named, with one contributor reaching for Respeecher when fidelity matters more than speed. Suno was named by only three, but all three are enthusiastic, and one argues music generation is the most underused lever in the field: a short can take an entirely different dimension with the right audio. The cut has not changed at all. Premiere, Resolve and CapCut do the assembly and finishing, because the AI clips still need real micro-trimming on a real timeline.

Ten of eleven contributors name Seedance 2.0, and when the panel describes their best shots they are usually describing Seedance output. But the real picture is a three-way verdict, and the camps are bigger than the headline suggests. Roughly six of the panel are Seedance-first. Three are Kling-first, and not as a compromise: they pick it for realism, motion consistency on action shots, and value. Two lean Veo 3.1. Two of those three engines are Chinese, the consensus pick among them, and they are iterating faster than the release cycle can track — the video frontier the panel works on is largely a contest between Chinese labs.

The practical rule from the panel: Seedance for cinematic motion, Kling for action, speed and budget, Veo when the face is the shot. If your film is built on character performance, the consensus pick is not your pick.

Just as telling is what fell off the list. Sora arrived on enormous hype and, in the words of one contributor, never delivered on the promise and is effectively fading out. Runway had its moment and has fallen behind on raw output quality. Google's Veo Omni push drew a similar shrug from three separate contributors, called repackaged rather than new. Underneath it all is a shared scepticism of any tool marketed on the idea that it replaces filmmaking, or that "Hollywood is cooked" — several flagged that messaging as a red flag, the sign of a product selling a future demo rather than something you can use today.

One finding we did not expect: content moderation is quietly functioning as a routing layer. One contributor defaults to Veo 3.1 and drops to Seedance or Kling specifically when Veo blocks her prompts, which for her work happens on legitimate scenes involving children and young people. Another avoids Seedance for over-blocking and runs the same shots through Kling instead. A third flagged that several tools will not generate indigenous people at all.

In practice the strictest filter loses the shot. Filmmakers do not abandon the scene, they reroute it to whichever model will render it, and the tool vendors who over-censor are donating usage to their competitors. For the platforms reading this: the panel is not asking for fewer safeguards on genuinely harmful content. They are asking for filters that can tell a school scene from something sinister.

Ask the panel what a beginner should use and what they themselves use, and a two-stage curve appears in the data.

Beyond the stack lies a third mode that only a handful of the most technical filmmakers operate in: the node-based pipeline. Four contributors described the same frontier. One builds entire productions in ComfyUI and calls it a professional environment rather than a generation tool, somewhere you stack workflows instead of starting from scratch every time, keeping around a terabyte of models locally for full independence. Another reaches for it whenever standard models fall short. A third says flatly that for character continuity, a node-based workflow is best. The trade is steep — a great deal of friction now for repeatability and control later — and most working pros never go this far. It is the edge of the field, not a rung everyone is expected to climb.

The advice for beginners follows directly: start in one app, and do not feel behind for staying there. Almost no one needs the pipeline. The panel moved from app to stack one job at a time, and only a few ever pushed to the frontier beyond it.

If you are already working

The most repeated advice has nothing to do with tools: spend your time in pre-production. Script, world, character design, visual language. The more defined the foundation, the faster and more consistent the generation. When you do prompt, think like a director and specify lens, lighting, movement and mood, because vague prompts give generic results. Build shots in layers rather than forcing everything into one giant prompt: shot, then mood, then camera, then character, then motion, then detail. Consistency comes from iteration, not prompt length. And let go of total control. Some of the best shots are the ones the model hands you that you would never have planned, and the filmmakers who get the most out of AI treat each model as a collaborator with its own personality, strengths and weird preferences, learning to communicate with each on its own terms.

Beyond mindset, the panel volunteered a set of specific hacks — small, repeatable moves that sharpen the work.

If you are starting out

Two things came up again and again. First, do images before video. Learn to generate and control a still image, get consistency, understand how prompting works, and only then animate. The panel is close to unanimous that text-to-video is the wrong place to start. Second, make short films and make a lot of them. A thirty to sixty second piece, shipped, then the next one. One contributor borrowed a line from film school: making one pair of shoes will not make you a cobbler, but making ten thousand will. The mistakes to avoid were just as consistent: do not chase every new tool instead of learning one pipeline well, and do not chase perfection on the first project or you will never finish it. And underneath everything, craft. Learn traditional filmmaking and photography, lighting, story structure, the show-don't-tell principle. The tools keep changing. Knowing how to tell a story is what lasts.

The panel agreed more on the problems than on almost anything else.

Long-form consistency is the number one gap by a distance. Holding a character, an environment and the emotional logic of a story steady across many scenes still takes a lot of manual work. The tool everyone wants does not exist yet: something that takes everything generated so far as context and uses it to make the next shot. As one contributor put it, an AI cinematographer, editor and VFX team all working from the same brain.

No single tool does the whole job. The pipeline is fragmented, and the constant handoff between four or five tools is itself the cost. If you are willing to give up control, one all-in-one tool can carry you. If you work professionally and want control, you are jumping between tools, and there is no way around it today.

Voice and emotional performance. Generated voices are fine for a rough cut and still fall short for finished work that needs real feeling. More than one contributor said anything serious still ends up using recorded audio or custom voice actors. One described the inconsistency precisely: one take is an unmistakably AI voice-over, the next is award-worthy, with no way to control which you get. The uncanny valley in performance is, for several of the panel, the single biggest thing holding the medium back.

Production readiness in motion. A shot can be high resolution and still not be usable. Skin that looks fine in a still feels artificial once it moves. Crowds and background characters dissolve into morphing. Night scenes are a recurring headache, with models inventing light and detail rather than letting a dark scene stay dark. As one contributor put it, darkness, silence and negative space are creative choices, and the tools treat them as problems to fix.

Cost. Credits surfaced everywhere: planning shots in advance to avoid regenerating fifteen times, upscaling bills forcing tool switches, generation scheduled around peak hours. The panel plans around price the way traditional productions plan around daylight.

The edges of the data point at where Issue 2 will look. One contributor is testing faster end-to-end movie workflows like Popcorn AI, and uses Runware as infrastructure to accelerate high-volume generation. The most advanced contributor benchmarks open and emerging video models, LTX and Wan among them, and is building toward LoRA-based video systems, custom-trained motion the way custom-trained characters already work in images. HeyGen earned the panel's only mention of avatar work and translation lip sync. And the local-model movement, running the whole pipeline on your own machine, is the quiet bet that the most technical filmmakers are already living in.

For all the tool talk, the panel kept circling back to the same idea. The people doing this well do not think of themselves as AI filmmakers. They think of themselves as filmmakers who use AI. The technology lowers the barrier and speeds up the work, but the taste, the rhythm and the point of view still have to come from a person. As one contributor put it as a warning to newcomers: do not build your identity around talking about AI, build work with it.

The most useful frame came from a filmmaker who has watched other creative fields go through this. When music went digital there was an uproar about the death of the craft. It did not stop anyone learning an instrument. What it did was let people make hits from a back bedroom. The gatekeepers changed, the floor for who could participate dropped, and audiences still decided what was any good. The bet on the panel is that AI filmmaking goes the same way. The barrier falls, more people get to make things, and the work that connects with an audience still has to be good. Cream rises.