Higgsfield published every prompt and asset behind Hell Grind, its 95-minute AI feature film, to launch a $1,000,000 global film festival.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
Higgsfield has published every prompt and asset behind Hell Grind, its 95-minute AI feature film, and made them free for anyone to read, copy, or remix. The film cost under $500,000 to produce, roughly $400,000 of which went straight to GPU compute. The release kicks off the Higgsfield Global Film Festival, a $1,000,000 short-film contest that opens for submissions on August 7, 2026.
That's the news. The interesting part is what's actually in those files — and one claim that needs correcting before anyone repeats it.
The prompts. Not the models.
This matters, because "open-source" gets thrown around loosely in AI. Higgsfield did not release weights the way Qwen 3.8 Max ships open weights. What they released is the full project canvas for three films — Hell Grind, Zephyr, and Mork — including every generated image, every video clip, and the complete text prompt behind each one.
| Film | Opens | What you get |
|---|---|---|
| Hell Grind | August 4, 2026 | Full canvas, all images and videos, complete prompts |
| Zephyr | August 6, 2026 | Full canvas, all assets, complete prompts |
| Mork | Announced Aug 3 | Prompts public |
Honestly? For most people this is more useful than weights would be. Nobody reading this is going to fine-tune a video model. But everybody making AI video is guessing at prompt structure, and now there's a 95-minute worked example sitting in the open.
Here's the number that stopped me: the prompts average around 3,000 words each.
Three thousand words. For a shot that comes out roughly 15 seconds long. That is not "cinematic wide shot of a demon, 4k, dramatic lighting." These are essays — full paragraphs reminding the model to respect physics, specifying lens choice and camera movement, locking character appearance across scenes, and explicitly telling the generator to avoid the plasticky "AI sheen" that makes generated video read as fake in half a second.
And the hit rate is brutal:
Sixty-four throwaway clips per keeper. If you've been feeling bad about your own reroll count, there's your benchmark. The professionals are burning through footage at a rate that would have gotten you fired on a real set.
That 64:1 figure is the most valuable thing Higgsfield published, and I suspect they didn't fully intend it as marketing. It quietly explains where the $400,000 went.
The stack is a mix of Higgsfield's own products and third-party models — which matters, because the marketing line has been that the film was "created entirely on the Higgsfield AI platform."
So it's a pipeline, not a button. Soul Cast handles character consistency, Cinema Studio handles shot assembly, and the heavy generation load gets split across Seedance and Veo. If you're building your own AI video workflow, that split is the pattern to copy — no single model does everything well yet.
No, and this is the claim that keeps getting repeated wrong.
Hell Grind screened in Cannes on May 16 and May 21, 2026. But it was not part of the official Festival de Cannes program. The screenings were third-party industry events happening in the city during the same window. When The Wall Street Journal reported that the film had debuted "at the festival," Cannes pushed back publicly, confirming to CineD that it was not in the official program and had instead been presented during an industry event organized by third parties.
The difference is not pedantic. "Premiered at Cannes" and "rented a room in Cannes during Cannes" are wildly different sentences, and only one of them is true. Higgsfield's own promotional framing has leaned hard on the ambiguity.
The coverage itself is real, though. Variety ran a full feature on watching the film, the WSJ covered it, and trade press including Screen Daily and The AV Club followed. I could not verify BBC News coverage, so I'd treat that one as unconfirmed until someone produces the link.
This is where Higgsfield is putting real money down. The prize pool splits across 14 winners:
| Prize | Amount | Winners |
|---|---|---|
| 1st Place | $500,000 | 1 |
| 2nd Place | $200,000 | 1 |
| 3rd Place | $100,000 | 1 |
| Audience Choice | $100,000 | 1 |
| Honorable Mention | $10,000 each | 10 |
| Total | $1,000,000 | 14 |
First place pays $500,000 — the same amount it cost to make Hell Grind itself. That symmetry is obviously deliberate, and it's a good pitch: win this and you've funded your own feature.
The jury has an actual heavyweight on it. Ed Catmull, Pixar co-founder, five-time Oscar winner, former president of Walt Disney Animation Studios, and a Turing Award laureate. Additional judges were being revealed through the week.
If you're a paying Higgsfield subscriber, yes. The rules as published:
Submissions open August 7 and close August 31 at 11:59 PM PT. Winners are announced at the end of September or early October. Judging runs in three stages: screening, shortlist, then jury verdict.
The subscription requirement is the catch, and it's the whole business model. A million dollars is a rounding error against the subscriber growth this will drive during August. Smart move, but call it what it is — this is a marketing spend with a film festival wrapped around it.
I have to be straight here: I haven't watched all 95 minutes, and neither has almost anyone else. Only roughly the first 22 minutes are publicly available. The film was also cut into an episodic series in 22 to 25 minute installments.
What I've seen of it — the trailer and the released opening — looks exactly like what it is. Character faces drift between shots. Motion has that weightless quality where nothing quite carries momentum. The action choreography reads as a series of impressive stills rather than continuous movement. Early online reaction has been largely negative, with critics calling the visuals artificial and the story weak, and I don't think that's unfair.
But judging Hell Grind as a movie kind of misses the point. Nobody's arguing it's a good film. The argument is that a 15-person team produced 95 minutes of watchable action-fantasy in two weeks for the cost of a single day of conventional feature shooting. Five years ago that was impossible at any price.
The plot, for the curious: four street thieves botch a heist, one gets pulled through a portal into the underworld, and the remaining three team up with a secret military unit to fight demon hordes across Tibet and feudal Japan. It's exactly as unhinged as it sounds.
If you're going to spend an hour with this release, spend it on the prompt structure rather than the films. Four things stand out:
Write long. The 3,000-word average isn't padding. Video models drift, and every sentence of specification is a constraint that reduces drift. Most people underprompt by an order of magnitude.
Name the physics explicitly. The prompts include direct reminders about gravity, momentum, and material behavior. Generators don't infer these — you have to say them.
Fight the sheen directly. There are explicit instructions telling the model to avoid the overly smooth, glossy look that marks generated video. Naming the failure mode in the prompt helps.
Budget for the reroll ratio. Plan for 60-plus generations per usable shot, not 3. Price your project accordingly.
That last one is the real lesson, and it applies well beyond video. The pattern shows up everywhere in AI production work right now — massive generation volume, aggressive filtering, small final output. It's the same shape as OpenAI's Astra running $2,000 of compute to land 10 math proofs. Generate wide, keep almost nothing.
If you make video: yes, unreservedly. A studio publishing its actual production prompts is rare, and these are detailed enough to learn from directly. Go read them before Higgsfield decides this was a mistake.
If you're just watching the AI video space: the interesting signal isn't the film's quality, it's the cost curve. $500,000 and 14 days for feature length. That number is going to keep falling, and the festival is Higgsfield betting that a lot of people are about to find out how far it's already fallen.
For more on where AI tooling is heading across other categories, see our roundup of AI tools replacing manual work in 2026.
Submissions close August 31. If you've been sitting on a short film idea and you already pay for Higgsfield, there are worse ways to spend three weeks than reading 3,000-word prompts and finding out how close you can get.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Claude Code 2.1.221 ships 39 CLI changes on August 4, 2026. Focus view is VS Code-only, and the real headline is a Bash permission-bypass fix.
Cursor's new Google Workspace plugins give agents access to Gmail, Drive, Calendar, Docs, and Sheets. What each one does, how to install them, and the risks.
Alibaba launched Qwen 3.8 Max — a 2.4T-parameter multimodal MoE model that just hit #4 on Frontend Code Arena with 1,668 points. At $2 per million tokens with open weights confirmed, it's the cheapest frontier-class coding model available.
