Comfy concluded the two-week open hackathon Open Source Higgsfield: the $25,000 prize pool was distributed among four projects, while ecosystem bonuses supplemented the picture with vendor API credits. Among the winners are a panorama generator, a Roblox game builder running on a local 27B model and a single RTX 5090, and a local movie studio with a cycle from film plan to video sequence with sound. All winning works are published on GitHub and can be installed in your own ComfyUI. The announcement page was updated on November 10, 2026.

image
image

What Happened

Results were finalized across four categories. Best Overall ($10,000) was awarded to the ComfyUI Panorama Stickers custom pack by nomadoor for creating panoramas. "Most Innovative" went to the Blocksfield project (Hcoder10): a text game pitch is transformed into a playable Roblox game by a local 27B model on a single RTX 5090. "Most Viral" was won by Imagination Studio (drowzeys, keys+Wasserman) — a local movie studio for Mac and NVIDIA DGX Spark, where the film plan is built using the Slate workflow, frames are rendered via Qwen-Image-2.1 and FLUX3-K, and video with sound is generated by MiniMax H3 and FastH3. "Open Weights Only" was taken by Paint a Plate from AInVFX — a ComfyUI pack for VFX where edits drawn on the first frame are transferred to the entire shot. Ecosystem bonuses: $2,000 for a VFX workflow on LTX (again Paint a Plate), $1,000 in Flux API credits for ComfyArt, $500 for Creativefield, $250 for genobi, and $1,000 in MiniMax credits.

Context

Open Source Higgsfield is an ecosystem contest for the ComfyUI platform, where generative pipelines are assembled from nodes and custom packs; the hackathon lasted two weeks. For Comfy, such a contest is a showcase of what end-to-end scenarios are already being built on top of specific models (FLUX3, Qwen-Image, MiniMax H3, LTX) and consumer hardware (RTX 5090, NVIDIA DGX Spark), from VFX retouching to Roblox world generation. A commercial mechanism is also noticeable: model vendors sponsor the ecosystem with API credits — from Flux to MiniMax H3 — receiving in return presence in open pipelines, while Comfy itself secures its role as a distribution channel for open workflows.

Why This Matters for the Industry

For builders, this is a catalog of reference pipelines, not a contest for the sake of a contest. Imagination Studio demonstrates an architectural pattern of multi-stage orchestration: film plan → frames via Qwen-Image-2.1 and FLUX3-K → video with sound via MiniMax H3 and FastH3 → voices and editing, all on consumer hardware. Blocksfield shows that a local 27B model is capable of driving an agentic cycle of "pitch → playable game" on a single RTX 5090, while Paint a Plate targets the unresolved task of consistency in frame-by-frame edits, transferring a drawn stroke to the entire shot. These demos set usage scenarios for node and model developers to orient themselves toward; ecosystem announcements also show a shift where hackathon victories serve as a substitute for formal benchmarks.

Why This Matters for Users

Practical value begins immediately: the winning repositories nomadoor/ComfyUI-Panorama-Stickers, Hcoder10/blocksfield, drowzeys/Imagination-Studio-by-keys-Wasserman, and AInVFX/ainvfx-paintaplate are open on GitHub and can be installed in your own ComfyUI in minutes. In the code, you can analyze ready-made "pitch → game" and "film plan → render" pipelines and transfer their patterns to your own tasks — from panoramas to VFX retouching. The limitation is hardware: Blocksfield is designed for an RTX 5090 with a local 27B model, while Imagination Studio is designed for a Mac and NVIDIA DGX Spark combination. And this is hackathon code: before responsible use, it is advisable to conduct your own quality and speed measurements rather than relying on the author's demo.

What Is Unknown / Limitations

The sources do not provide an evaluation methodology: no metrics, no comparison with baselines, and no published jury criteria. The "Most Viral" category measures audience attention, not quality, while "Open Weights Only" fixes a limitation on model weights, not the technical level of the solution. The strongest claim, Blocksfield, remains a single demo: the frequency of successful launches and the quality of the generated game code have not been evaluated. Paint a Plate has neither a method description nor a numerical evaluation of consistency, so transferring an edit to the entire shot is a claimed capability, not a verified result. Finally, only the winners of the two-week hackathon are publicly visible: the sample is biased, and "it worked for the author" is not reproducibility.

Sources

Author

Look at AI, editorial team