#OpenSourceHiggsfield final day: six open studios to try

OpenField is the standout. wexloops' MIT app puts a screen on top of ComfyUI workflows for video, stills and 3D, all on open weights, and its launch post passed 230 bookmarks overnight. Per the README it has a Director panel (camera, lens, film stock, grade, an era scrubber and speed ramps compile into the shot), Scene (a local LLM plans 2 to 6 shots and cuts them together), Stage (block a shot with boxes, figures or your own .glb models plus a camera path, rendered through the LTX-2.5 layout IC-LoRA), Mirror (film a move on your phone via a QR code and put it on a character), Reshoot (re-aim the camera on an existing clip), seamless loops with a seam score, HDR finishing to ACEScg EXR and HLG masters, and 31 MCP tools so Claude Code can drive every app. Every take opens as the exact ComfyUI workflow that made it. It wants an NVIDIA card with 24 GB for video (built on a 32 GB RTX 5090); 12 to 16 GB runs stills only.

Kitty Open Studio (Druid Cat, MIT) is the timeline option: 33 curated open-weights workflows as cards, a cut editor, and an MCP connection so Claude or ChatGPT can direct a whole music video. It's set up to run on a Runpod pod, pay-per-second serverless or your own GPU. OpenKatana (Apache-2.0, on the Comfy Registry) goes the other way: an editing engine your agent drives, where photos, footage and music go in and a cut with transitions, motion type and music comes out. Per its FAQ the standard editing needs no dedicated GPU.

Cheapsfield (Apache-2.0) is a Windows desktop studio for product and architecture work: sketch to packshot or render, relighting, virtual staging and 1080p video with sound from LTX 2.5, using only core ComfyUI nodes and no paid APIs. It lists 12 GB of VRAM as the minimum and 16 GB as recommended. OpenIllusion (MIT) turns any ComfyUI workflow into a one-click template and shows MiniMax H3 running on a 16 GB card in its demo; it can also call cloud models through Comfy Router, billed to your own Comfy API key. Its templates include character swaps, so only swap people who've agreed to it.

Rapid-H3 is for one machine. The Apache-2.0 node pack claims MiniMax H3 clips with audio in about 117 seconds on an NVIDIA DGX Spark, against 590 to 630 seconds for stock ComfyUI, using custom INT8 kernels, six steps instead of ten and a draft-then-refine pass. Those are the author's numbers on their own Spark. Its README also flags that the MiniMax H3 community licence excludes the USA, EU, UK and South Korea unless you apply separately, which matters for every H3 project on this list.

Our take: OpenField is the one to install if you have a 24 GB card, because Stage and Mirror cover the blocking and performance control most local setups lack. Kitty Open Studio and OpenKatana are the better fit if you'd rather direct through an agent and a timeline. Check each model's licence (LTX-2.x is free for companies under $10M a year in revenue, per OpenField's notes) before client work.