102 open-source AI video studios and a legal threat: what Comfy's contest delivered

What this is about

Over the past two weeks we've covered IV Studio, Mago Studio, Openjutsu, Reshoot 4D and H3MD Director one by one. They're all free programs built on ComfyUI, the interface where you assemble AI workflows from connected building blocks ("nodes"). Several of them come from the same campaign: Comfy Org, the company behind ComfyUI, called on the community to rebuild the features of the commercial video platform Higgsfield using open models. That produced more than 100 projects. Higgsfield responded with a lawyer's letter.

We explain what's behind it. We also run the numbers on the hardware these projects need, how open the favored video model MiniMax H3 really is, and what you can realistically use on your own machine today.

A contest with shifting dates

According to the ComfyUI Wiki, the contest started in early October as the "Open Call: Comfy Dev Platform Challenge" with a submission deadline of October 19. From October 15 it continued under the punchier name #OpenSourceHiggsfield. The rules: entries had to be working apps with a user interface, not just exported workflows. They had to build on Comfy's developer platform (API, Router or SDK) and be published on GitHub under an open-source license, with installation instructions and a demo video of no more than two minutes. Partners were NVIDIA, Runpod, Lightricks (LTX), Black Forest Labs (FLUX) and MiniMax.

Do the math: $12,000 is paid out in cash, plus $5,250 in credits and three graphics cards. You only reach the advertised $25,000 if you value each RTX 5090 at around $2,600.

The dates don't line up either. On October 29, Comfy posted on X that the deadline was October 30 and winners would be announced on November 6. The official contest page now lists November 3 (9 a.m. Pacific) as the deadline, an extra build night in New York on November 2, and November 10 for the announcement. Comfy doesn't explain when or why the dates moved.

The dispute: the name, not the code

On October 29, Comfy co-founder Yoland Yan posted a three-page letter from Higgsfield on X. According to a summary by Art Realm AI, it's dated October 28 and signed by the company's Associate General Counsel. Higgsfield objects to the name "HIGGSFIELD" appearing in the title, hashtag, URL and submission form, and to CEO Alex Mashrabov appearing in the campaign video. That, the letter says, creates the false impression that Higgsfield is behind the contest. It cites a US trademark and the Lanham Act, the US trademark law. The demands: stop using the name, remove the video, preserve records, and confirm all of this in writing within two days. One sentence in the letter stands out:

Higgsfield has no objection to open-source software, competition, or commentary.

Yan framed it as a demand to shut down the entire challenge within 48 hours. Japanese blogger dskjal thinks that's overblown: the letter mainly asks Comfy to stop using the name and take the CEO out of the video. Comfy's response was defiant:

The challenge stays up.

We found no public response from Higgsfield. Comfy's justification is also shaky: it claims Higgsfield had promised to open up its platform. ExplainX found no evidence of that. All we found was the reference app "open-higgsfield," released at the end of September — a Next.js front end that only works with Higgsfield's own paid API. For context: Higgsfield was valued at $5.4 billion in August and says it has 32 million users.

What came out of it: 102 projects, more LTX than H3

The list maintained by Jutsu Studio counts 102 contest entries. Most are all-in-one studios (26), followed by camera and reshoot tools (15) and character swapping in the style of Higgsfield's "Genjutsu" (14). By far the most popular is Openjutsu: according to one list it had 656 GitHub stars on October 31, 910 on November 1, and more than 1,600 today. Next come ComfyArt (around 400) and ComfyUI-Continuity (around 140). Most other projects have only a handful of stars. Some ship without any license file at all, OpenField for example. Legally, that means you're not allowed to reuse them.

Of the five projects we covered individually, three are confirmed contest entries: Openjutsu, IV Studio and Reshoot 4D. H3MD Director doesn't mention the challenge in its README, and we found no evidence either way for Mago Studio. The pattern is still clear: within two weeks, dozens of front ends appeared for the same open models.

Counting through the lists turns up a surprise. Where a project names its video model, it's LTX (version 2, 2.3 or 2.5) in about 13 cases and MiniMax H3 in only 4 to 5. Many entries simply say "open weights." Our take: the only cash bonus went to LTX projects ($2,000), while H3 only got API credits. LTX also runs on smaller cards. Both probably tipped the balance.

MiniMax H3: the Stable Diffusion of video?

Outside the contest, H3 has nonetheless become the foundation of an ecosystem. MiniMax released the weights on August 3: a dense 33-billion-parameter transformer that generates video and stereo audio together. Three months later, Hugging Face shows 209 finetunes, 188 adapters, 86 quantizations and 11 merges. Quantization means shrinking a model by storing its numbers at lower precision. The version repackaged by Comfy racked up around 31.9 million downloads last month. That counts file downloads, not users. The last time a single model sparked a comparable surge was Stable Diffusion for images.

But not everything is open. H3-Context-IR, which preprocesses prompts, remains a hosted service according to MiniMax. H3-Regenerate-2K, the 2K re-render, has "not yet" been released, and the fast sparse attention is missing from the first version. The community is now filling these gaps itself. Open-H3-Regenerate-2K upscales a 768p draft to 2560×1440 in three steps. But it was only benchmarked on a B200 data-center GPU (22 to 87 seconds per 5-second clip), and there are no VRAM figures.

How much VRAM the studios really need

The projects' figures vary widely. All of them come from the developers themselves:

Independent user reports in a Hugging Face discussion temper expectations. On an RTX 3060 with 12 GB, 3 seconds of video at 1376×768 with 24 steps took about 20 minutes. With a Turbo LoRA (an add-on module for speed) at 864×480, 5 seconds still took 4.5 minutes. A 4060 Ti with 16 GB needed 62 minutes for a 22-second clip. Even on an RTX 5090 with 32 GB, one user hit out-of-memory errors at around 35 seconds of video. A writer on Dev.to calls running it on 12 GB "possible, but not comfortable" — more for getting to know the model than for real work.

The widely shared claim that RefMods keeps characters and voices consistent on a 6 GB card comes from a single Reddit post using a stack of memory-saving nodes. The RefMods repository itself gives no VRAM figures. It describes itself as "work in progress" and lists voice transfer as a known bug. The Reddit author, by contrast, reports consistent voices. Those two accounts contradict each other.

The license trap

For readers in Germany and the rest of the EU, the MiniMax H3 Community License is the most important point. According to the license text on Hugging Face, the EU, the UK, South Korea and the US are excluded from its territory. Using the model or its outputs outside that territory is not permitted. Other terms:

• Companies with more than $20 million in annual revenue need written permission.
• Commercial products must display "MiniMax H3" prominently in the interface.
• Outputs may not be used to train other AI models.
• Hong Kong law applies.

The license covers all the derivatives listed above, from the Pika adapter to the 2K replacement.

Ironically, Comfy offers a way out: the company is an official reseller of H3 licenses. The Professional tier starts at $5,000 a month and includes around 46,250 seconds of video. LoRA training is included. Excluded are resale, model marketplaces and "inference-as-a-service." Comfy doesn't say whether a hosted service with your own LoRAs falls under that. On top of that, some building blocks carry their own restrictions: InsightFace, which IV Studio and Openjutsu use for face matching, is licensed only for non-commercial research. According to Kinoloom, Qwen Image 2.1 is under a non-commercial research license.

Our take: MiniMax being a contest partner while Comfy sells H3 licenses isn't a scandal. But it shows that "Open Source Higgsfield" is also a sales campaign.

What you can actually run locally today

• With 8 GB: Kinoloom is explicitly tested on this, but then relies mainly on LTX-2.5 and image models. For H3, you're left with GGUF quantizations and a lot of patience.
• With 16 GB and 64 GB RAM: Openjutsu (about 6 to 13 minutes for a 6-second clip at 480p), IV Studio with smaller previews, Backlot. On Windows, these are the most accessible starting points.
• With 24 GB: Reshoot 4D, IV Studio at full quality, H3MD Director for longer videos.
• Without a powerful card: IV Studio and Mago Studio can render through your own Runpod, Modal or Vast.ai account. IV Studio estimates around $0.13 for a 10-second shot on an L40S.
• Mac: essentially not supported. Almost every project requires an Nvidia GPU.

The bigger picture

The contest showed how quickly paid features can be rebuilt with open models. There are now usable free alternatives for character swapping, camera reshoots and short-film assistants. So far, though, almost all descriptions of quality come from the developers themselves. Many projects are only a few weeks old, have a handful of stars, and only run after you've set up 20 or more node packages. Usability is no longer the biggest problem — hardware and wait times are.

Our take: the comparison with Stable Diffusion holds for the ecosystem — adapters, quantizations and tools. It doesn't hold for openness. Stable Diffusion could be used worldwide; H3 excludes the EU and withholds key parts like the 2K upscaler and the prompt system. If you want to work legally and commercially in Germany or elsewhere in the EU, LTX-based projects are currently the better choice. The trademark dispute with Higgsfield, on the other hand, is unlikely to change much: it targets the name, not the code. According to Comfy's contest page, the winners won't be announced until November 10.