Challenge One

Dos Bros — Open Weights Film Challenge

Challenge One · Rules & Eligibility Short films, open models, and whatever card you've already got.

0. What This Is, In One Paragraph

Nobody buys their way past someone more creative. Every entry runs on open weights, on hardware you could actually buy, in a weight class against comparable cards. We hope you'll share how you did it, because that's how everyone gets better. Run by two brothers with day jobs, funded out of pocket, free to enter, forever.

1. The Four We're Not Planning to Bend

Almost everything in this document is our best guess as of today, stamped Challenge One. We expect to get plenty of it wrong and fix it as we go.

These four are the ones we don't intend to move, because the whole thing stops meaning much without them:

No money from anyone with a stake. Prize funds don't come from a company with a commercial interest in the eligible models or the eligible hardware. Individuals can chip in. Model vendors, GPU manufacturers and cloud providers can't.
The organizers don't enter. Not under a handle, not as a team member, not in an edition we run.
No judge has a commercial stake in the eligible models or hardware, and no judge can enter the edition they judge.
Model choice carries zero weight in judging.

If we break one of these, the thing has stopped being what it was for. Hold us to it.

Everything below this line is Challenge One, written in good faith, and open to revision the moment somebody shows us a better way.

2. Weight Classes — The Reason This Exists

An 8GB card and a 32GB card are not the same sport.

Class	VRAM
Featherweight	8GB and under
Middleweight	9–16GB
Heavyweight	17GB and above

Declare your class at entry; render logs confirm it.

The Grand Prize is judged across all classes. A Featherweight entry can and should win it. Getting coherent character work out of 8GB on GGUF quants is a harder problem than brute-forcing it on a 5090, and it should not lose to one.

Apple Silicon — open question for Challenge One. Unified memory breaks the bracket: a 64GB Mac has more addressable memory than a 4090 and far less compute. Until we have data, Apple Silicon entries compete in Heavyweight and are marked as such. Tell us if that's wrong — this is exactly the kind of rule we expect to revise.

We do not require step count or quantization disclosure. The research doesn't support treating them as quality proxies, and requiring them would imply a hierarchy that isn't there.

3. Open Weights — The Eligibility Rule

Every AI-generated frame must come from a model whose weights are freely downloadable by any member of the public.

This excludes Veo, Kling, Sora, Runway, Luma, Hailuo, Vidu, Seedance, Midjourney, and every other closed or API-only model.

That exclusion is the point, and it isn't a shot at them. Those tools are good, and sending a job to a datacenter genuinely gets you more power than a card in your desk — pretending otherwise would be silly. We're not competing with them and we're not trying to show them up.

We're after a different question: what's actually possible on the machine you already have? Plenty, we think, and hardly any of it gets shown off properly, because the showcase usually goes to whatever had the biggest render farm behind it.

Open source thrives on demonstration. People pick up the free tools when they see somebody do something great with them, not when someone argues they ought to. So this isn't a protest — it's a focus. Everything here runs on tools anyone can download and use for nothing, and the point is to find out how far that goes.

And it must run on consumer hardware

Not permitted: any hosted API or generation service (including services that run open models for you), and any datacenter GPU — H100, H200, A100, B200, MI300, TPUs, or equivalent. Renting datacenter compute is out.

Permitted: consumer graphics cards — NVIDIA RTX 20/30/40/50 series, AMD Radeon consumer cards, Intel Arc, Apple Silicon. Your machine, a friend's machine, a machine at work. We're not checking the deed. We're checking the card.

Render logs print the GPU. An H100 cannot pretend to be a 3060. This is the version of the rule we can actually enforce.

Approved model list — Challenge One

Enumerated, not defined. "Is it on the list" is a question you can answer in ten seconds; "is it open source enough" is an argument nobody wins.

Video: Wan 2.1 / 2.2 · LTX-Video / LTX-2.x · HunyuanVideo · CogVideoX · Mochi 1 · Open-Sora · SCAIL / SCAIL-2 · SkyReels · AnimateDiff · Wan-Dancer-14B · MiniMax H3

Image: Stable Diffusion 1.5 / SDXL · Flux · FLUX.2 · Qwen-Image · Krea 2 · Illustrious · Mage-Flow · Ideogram 4

Audio / voice: any model with downloadable weights.

LoRAs, fine-tunes, quantizations (GGUF, FP8, INT8, NVFP4, MXFP8), and community checkpoints built on the above are all permitted.

To add a model: open a request before the deadline. The only test is whether the weights are publicly downloadable from the official org — verified against that org's Hugging Face page, not a third-party claim.

This list is revised every edition. Models ship weekly; a list written in August is stale by October.

Every frame is AI-generated

The picture is the AI's. Every frame of the film itself should come from an open-weight model running on your machine — not live-action footage, not stock, not photography, not hand-drawn cels, not a 3D render passed off as a shot.

That's stricter than most competitions, and it's deliberate: the question we're asking is what can these models do, and the answer gets muddier every time a real camera helps out.

Not counted as frames, and completely up to you:

Titles, credits, lower thirds, subtitles. Models are still bad at text, and nobody should lose a competition to a wobbly title card.
Transitions — cuts, dissolves, wipes, whatever your editor does between shots.
Colour, grading, compositing, VFX, motion graphics applied on top of AI frames.
Conventional (non-generative) upscaling and frame interpolation.
Yourself, in your own footage, if you want to top-and-tail with a piece to camera. Say so in the attestation.

Audio is yours to do whatever you like with

No restriction at all. Generate it with open models, score it yourself, hire a friend, license a track, record dialogue in a cupboard. Music and sound design aren't what we're testing.

One practical warning, and it matters more than it sounds: entries live on your YouTube channel. Copyrighted music will get your video muted, blocked, or claimed, and a blocked entry can't be watched or judged. Use music you made, music you've licensed, or music generated with an open model. We'd hate to lose a good film to a Content ID strike.

Edit however you want

Resolve, Premiere, Kdenlive, CapCut, anything. The rule is about generation, not about your editor, your workflow, or how much work you did by hand afterwards.

4. How a Round Starts

Honestly? We don't know if this will work.

We're two brothers with day jobs trying something that doesn't exist yet, and the whole thing rests on a bet — that there are enough people out there making films on a card they actually own who'd want somewhere to put them. We think there are. We might be wrong.

So we're not announcing a deadline and hoping anybody turns up.

Sign up, and we'll email you when there are enough of us

That's the whole mechanic. Tell us your weight class and one line about what you'd make. It's free and it commits you to nothing.

When enough people are in, everybody gets the same email on the same day: start date, deadline, and the signature element. Then the clock runs and it's a real deadline.

The list of who's in is public — handles, not just a number. So you can see who else is here, whether the count is real, and whether you're in good company.

And if it turns out there aren't enough of us, you'll get an email saying that too. We're not going to leave a page sitting here for eight months advertising a start date that never came. If this doesn't fly, we'll tell you, say thanks, and leave the archive up.

The prize is already set aside. Not pledged, not "subject to sponsorship" — it's sitting in an account waiting for someone to win it.

This is an indie thing run by two people in their evenings. It might stay small. It might not be what you were hoping for, and if so, fair enough — genuinely, no hard feelings. But we think it'll be fun, the money's real, and nobody here can outspend you.

5. Submission

Runtime: 90 seconds to 3 minutes, including titles. One range for every class.

Ninety seconds is roughly 6–18 separate generations, which means continuity, character consistency, pacing and a real edit — the craft this is actually about, and more than any single prompt can fake. The range is wide on purpose: if you're render-bound on a small card, 90 seconds is a complete entry and eligible for everything including the Grand Prize. If you've got the hours, use all three minutes. We're not setting different targets per class — a weight class decides who you're matched against, not how much is asked of you. A Featherweight entry that hits three minutes should be able to exist, and be as impressive as it sounds.

Resolution: 720p minimum for Featherweight, 1080p minimum for Middleweight and Heavyweight. Conventional upscaling is permitted and does not count against you — the rule exists so the class we built to protect isn't penalized by a spec its hardware can't natively hit.
Where it lives: YouTube, public. Upload it to your own channel and send us the link. We embed; we don't host, store, or re-upload anything. It stays your video on your channel, with your views and your comments.
Tag it #DosBrosChallenge in the video description.

Two reasons, both practical. It makes every entry findable in one search — by us, by the judges, and by anyone who wants to watch the whole field without waiting for the showcase. And the upload date quietly corroborates that the film was made during the round, which saves us asking you for anything more intrusive.

It has to be public rather than unlisted, because unlisted videos don't surface in search and the tag stops doing its job. If that's a problem for you, tell us — we'd rather hear it than lose the entry. - Limit: 3 entries per person or team. - Language: any; non-English requires English subtitles. - Age: 18+. Minors can't enter binding agreements and we can't pay them. - Fee: none.

Teams are fine. Name everyone. Prize goes to the submitting entrant to divide.

Made for this round — the originality rule

Your entry must be made during the round, for this round.

Created on or after the round's start date. Not older work. Not a re-edit, re-cut, or re-score of older work. Not a longer piece trimmed down.
The signature element. When the clock starts we announce a required element — an object, a word, a line, a specific shot. It must appear in your film. This is how timed film races have proven originality for decades: you can't retrofit it into finished footage without it showing.
The work must be yours. Your own generation, your own edit. Submitting someone else's film is the one thing here we'd consider genuinely disqualifying beyond this competition.

You may post your entry anywhere you like, whenever you like. We don't require a premiere and we don't care if it's already on your channel — only that you made it during the round. Share it while the round is open if you want to; that's your work and your audience.

You may enter it in other competitions too. We're not asking for exclusivity. If it also fits somebody else's open call, go.

How this is actually checked: mostly by §7. To claim a prize you publish render logs, model hashes, seeds, and a workflow that has to reproduce on the card you declared. Render logs are timestamped and model versions are dated — you can't have used a model that didn't exist yet. In release-triggered rounds, the model itself does most of the work: if the challenge is built around something that shipped last week, older work isn't disallowed so much as impossible.

Realistically we spot-check winners. Nobody fakes a render history for $500.

6. Content Standard

Written in plain language rather than borrowing a national rating system, so it reads the same everywhere.

The test is simple: could we screen it publicly and put it on the front page? If yes, you're fine.

Permitted, and we'd hate for anyone to self-censor over this: attractive characters, romance, sensuality, action, tension, peril, stylised violence, dark and difficult themes handled with craft. Films are allowed to be films.

Not permitted: - Explicit sexual content or nudity - Graphic gore, torture, or realistic depictions of severe injury - Any sexual or exploitative depiction of a minor — instant permanent ban, reported - Hate content targeting protected groups - Content promoting self-harm

If you're genuinely unsure where something sits, ask us before you spend a fortnight on it. We'd far rather have that conversation early than turn away something good.

Nothing auto-publishes. Every entry is reviewed by a human before it becomes visible. That's what makes the next rule enforceable rather than aspirational.

The likeness rule — the one we won't budge on

No real people, living or dead.

No celebrity LoRAs, no face swaps of real individuals, no recognizable public figures, no likenesses of anyone who has not given written consent. Entries containing them are disqualified, without appeal or resubmission.

This is stricter than most competitions and it is deliberate. The open-model ecosystem carries a real and well-documented reputation problem around non-consensual likenesses. This competition exists to show what the open ecosystem is for. One bad entry becomes the whole story. We are not risking it.

Original characters, fictional beings, animals, and stylized humans are all fine. Yourself, with your own footage, is fine.

7. Show Your Work — Please, But We're Not Making You

Publishing your workflow isn't required — not to enter, and not to win.

We thought hard about making it the price of a prize, and decided against it. Open source didn't get here by forcing anybody to share — it got here because people chose to, over and over, for no reward beyond someone else being able to build on it. A rule that extracts your method at gunpoint isn't in that spirit, and honestly it would just make people enter with their second-best work.

So it's a genuine choice, and we've tried to make the generous one worth taking.

What we do need — attestation only

Every entry declares:

every generation model used (from the approved list)
LoRAs / fine-tunes used
a one-line description of your pipeline
GPU model and VRAM, and your declared weight class
confirmation that no closed model or hosted service generated any frame, and that all generation ran on consumer hardware

That's the whole requirement. It's how the weight classes and the open-weights rule are enforceable at all, and it takes about a minute.

What we'd love — and how we say thank you

If you're willing, publish the method alongside your film:

Workflow artifact — ComfyUI JSON, Wan2GP settings export, a script, a CLI invocation, or a written parameter list. Interface-agnostic. The test is "can someone reproduce this," not "did you use the right UI."
Prompts — positive and negative, per shot or scene
Model files and hashes, with links
Seeds and key sampler settings, where you have them
A short write-up — what worked, what broke, what you'd do differently

Entries that share get:

a Shared mark next to them in the showcase, permanently
inclusion in the workflow archive, organised by model so it's actually findable in a year
eligibility for Most Useful Workflow, a community-voted award decided after the archive has been public a few weeks — people vote for what they actually downloaded and ran
our genuine thanks, which is worth what it's worth

Partial is welcome. Just the prompts. Just the seed. Just a paragraph on the one thing that finally worked. A single sentence saying "the trick was two-pass VACE at 0.6 denoise" helps somebody. You don't have to hand over the whole graph to be generous.

You keep full copyright in your film either way. Sharing the method is not sharing the work.

Why we care: right now, posting a workflow is pure cost — you give away your edge and get some upvotes. That's why people post the video and not the JSON, and why the good stuff lives in Discord threads that are unsearchable a month later. We can't fix that with a rule. We can make it a bit more rewarding, and say thank you loudly.

8. Judging

A panel of named judges plus a separate community vote.

Named before signups open. Every judge is a real, credited person. If you can't see who's judging, don't enter — ours or anyone's.

Anyone can put themselves forward. We'd rather the panel came from people who actually use these tools than from names we borrowed, so there's no invitation-only tier — tell us you're interested, send a couple of links and a line about what you make, and that's the application.

Panel composition: - Open applications, judged on evidence: your work, your published workflows, nodes you've built - We may also invite people directly; either route lands in the same seat - If you think there's a better way to run a panel, tell us. We've made a lot of this up and we'd rather hear a good idea early than defend a bad one later - Odd total. Community seats rotate each edition. - Judges can't enter the edition they judge — but past that, do as you like: judge one and enter the next, judge every one, or dip in when it suits. We're not precious about it. - Entries anonymized before panel review where feasible.

Neither organizer enters — ever. One of us sits on the panel as a single voice among an odd number, never as the decider. The other doesn't. We publish the reasoning behind every result, so you can argue with the argument rather than guess at it.

Awards

Grand Prize — best entry overall, any weight class
Class Champion ×3 — best Featherweight, Middleweight, Heavyweight
Most Useful Workflow — community-voted, decided after the archive has been public for a few weeks. People vote for what they actually downloaded and ran. Usefulness is measurable by adoption; elegance isn't.
Craft awards (whole-pool eligible, no field splitting): Best Character Consistency · Best Motion · Best Explanation · Best Sound
All selected entries — official selection laurel, screening in the showcase, inclusion in the workflow archive

No weighted percentage rubric. Five people watching thirty short films are not computing a 76.4, and pretending otherwise is theater. Judges argue in plain language and we publish the reasoning.

We are explicitly not judging on polish parity with closed frontier models. A stylized piece with real character consistency beats a prettier set of disconnected shots. That's the whole thesis.

Model choice is recorded as a tag on every entry and carries zero weight in judging. It's distribution data for deciding whether future editions split into per-model tracks — not a preference.

The community vote is a separate award, not a tiebreaker. When the jury and the audience disagree, that's a feature, and we'll say so out loud.

9. Prizes & Payment

Grand Prize: $500 USD
Class Champions: laurel + featured placement + archive inclusion
All selected entries: laurel, screening, archive inclusion

The grand prize is judged across all classes. A Featherweight entry can and should win it.

Payment: Wise by default. USDC on request if that's easier where you are — PayPal is unreliable or unavailable in a lot of countries and our winner may well be outside the US or EU.

We never take crypto in. The prize pool is dollars, out of pocket. Paying a winner in a stablecoin because they asked is a payment rail, not a token project.

Every payout is published — date, amount, method — on the results page, permanently.

Winner is responsible for their own taxes. The prize is not contingent on publishing anything — see §7. We cannot pay into sanctioned jurisdictions; void where prohibited.

10. Disqualification

Closed-model generation anywhere in the entry
Non-AI footage used as frames (live action, stock, photography, hand-drawn, 3D) beyond the carve-outs in §3
Generation via hosted API or service, including services running open models
Datacenter GPU use, or misdeclaring your weight class
Real-person likeness without consent
Work created before the round opened, or a re-edit of older work
Missing the round's signature element
Submitting work you did not make
Content standard violations
Vote manipulation, bot accounts, sockpuppets
Misrepresenting your pipeline in the attestation

Spot-checks: we may privately request source files, intermediate frames, or render logs to confirm the attestation in §7 — that's the declared card and models, not your method, and it stays between us. Refusal is disqualification. Realistically we spot-check winners only — nobody fakes a render history for $500, and a published workflow has to actually reproduce on the claimed card. Largely self-correcting.

Appeals: 72 hours, one email, reviewed by a judge who wasn't on the original call. We owe everyone one clear explanation. Nobody is entitled to a debate.

11. Rights

You keep everything. Full copyright in your film stays with you.

By entering you grant permission to screen the work in the showcase, embed it on the site, and use stills and clips to promote the challenge, always with credit.

If you choose to publish workflow materials (§7), those are licensed CC BY 4.0 so others can build on them. That applies only to what you decide to share, and never to the film itself.

12. Who Runs This

Dos Bros — two brothers who make animations, and have day jobs.

We've been in this scene a while: thousands of renders between us, all of it local, on cards originally bought for gaming. We kept seeing genuinely good work scroll past and disappear under a dozen upvotes, and eventually got tired of waiting for somebody else to build an occasion around it. So this is that.

On the prize: $500 is small on purpose. Small enough that nobody's faking a render log to win it, small enough that this stays about the films — and still enough to matter if you're building something on your own. We'd run it for nothing; we just figured that if we're asking for a few weekends of your time, we should put something on the table to start it off.

Self-funded. The prize is our own money.
Neither of us can enter, in any edition.
One of us judges, as one vote on an odd-numbered panel. Never as the sole decider.
We use these models ourselves, which is exactly why model choice carries no judging weight.
No sponsor, no vendor, nobody with a product in the race.

We are not a marketing team and we're not going to pretend to be one. This runs when we can run it well, and if we ever stop, we'll say so and leave the archive up on GitHub rather than let a dead page rot on the internet.

13. Feedback — And What We'll Do With It

After every round we publish a retrospective: what the payouts were, what people told us, and what we're changing. Including the criticism. Nothing gets deleted and everything gets a response.

We can promise that credibly because we're small. Fifteen entrants is easy to actually listen to.

14. There's Another One, And You Should Know About It

The Arca Gidan Prize — run with Banodoco, ComfyUI and Lightricks — is an open-source AI art competition, two editions in, with a prize pool around $65,000 and roughly 90 artists per edition. Their payouts are verifiable on-chain. It's good and it's bigger than this.

Here's the difference, stated plainly:

	Arca Gidan	Dos Bros
Open models	≥75%	100%, no exceptions
Hardware	Unrestricted — "the requirement is on the models, not the tools"	Consumer cards only, no datacenter GPUs
Weight classes	None	Three, by VRAM
Funding	Partners incl. a model vendor	Self-funded, no sponsors ever
Prize	~$65,000	$500

If a bigger prize is what you want, go enter theirs — genuinely. This one exists for a different reason: to find out what people do when nobody can out-spend them.

Why This Exists

Of the AI film competitions running today, nearly all accept any tool. The ones that restrict, restrict to a closed platform. One restricts to open models — and still lets you rent a datacenter to run them.

Nobody runs one where the hardware is the constraint. Not because it isn't wanted, but because competitions are funded by vendors buying distribution, and no vendor wants a rule that says you can't buy your way in.

So this one is small, self-funded, and run by people who actually use the tools.

It's built for the tinkerers — the ones with a gaming PC, a ComfyUI graph held together with custom nodes, no ad budget, and no subscription to a frontier model. People making genuinely interesting work who currently have nowhere to put it where they aren't outgunned before they start.

Everything here follows from that. Open weights, because you shouldn't have to pay to generate. Consumer hardware, because you shouldn't be able to buy your way past someone more creative. Weight classes, because 8GB and 32GB are not the same sport. Published workflows, because the next person shouldn't have to rediscover what you figured out.

If that sounds like your kind of thing, it probably is.

Get on the list