The plan and the build record

Higgsfield × ComfyUI — what's being built

Prepared: 2026-10-29 · v2 (corrected after the competition-brief lane landed) Event: Comfy Org — Open Source Higgsfield · LA Build Night, in partnership with AI On The Lot Submission deadline: 2026-10-30, 09:00 PT — ≈15.5 hours from this document Status: the breakdown the operator asked for, before building starts.

v2 changes the picture in four places. The event's own premise is false; the "four inspiration prompts" are three cards; the licence landscape contains a trap that disqualifies the obvious build; and the obvious rebuilds are already taken. §1 and §5 are rewritten. The entry thesis survives all four — and is sharper for it.

0. The 60-second version

Comfy Org is running a competition inviting the community to rebuild Higgsfield's most popular apps with ComfyUI. Three app cards are named: Cinematic Studio, Re-shoot a video, Openjutsu.

We are entering with the one thing nobody else has: a cinematographer that measures.

Magieva has already built, in production Go, a 38-rule machine-applicable cinematography rulebook where each rule returns a measured value, a pass/fail, and the camera lever that fixes it — plus a 24-iteration geometric camera solver, a sequence-level continuity engine (the 180° line, the 30° rule), and a metric 3D scene layer with real optics.

Ported into ComfyUI as a custom node pack, that is a genuinely novel contribution to the ecosystem — which is precisely how the Most Innovative category is defined ("novel custom node packs or uses of ComfyUI").

One-line thesis: Every other re-shoot app asks you to aim a camera and hope. Ours tells you the frame is wrong, by how much, and which way to move — before you spend the render.

1. What Comfy Org is actually asking for — corrected

1.1 The premise is false

The event page says "Higgsfield said they'd open source their platform. Let's help." Higgsfield never open-sourced its platform, and no source shows it ever promised to. The September 2026 viral claim conflated higgsfield-ai/higgsfield — an Apache-2.0 LLM-training/GPU-orchestration framework, now archived — with the commercial studio. Higgsfield's eleven public repos are thin API clients or unrelated. The "open source Higgsfield studio" repos dominating search are third-party.

This matters for positioning, not for legality: there is nothing to fork. Cinema Studio 4.0 runs on Seedance 2.5 (ByteDance) per Higgsfield's own help centre; every other Higgsfield-branded model is closed with unpublished architecture. A rebuild is a from-scratch reimplementation of the product surface on open models.

1.2 The "four inspiration prompts" are three cards — and they are Comfy's own apps

The event page presents three app cards, each a Comfy-Org MIT reference app, not a Higgsfield product:

There is no fourth app. There is no Higgsfield app called "Re-shoot" (that is an unrelated iOS app). The event page's own copy calls these "inspiration prompts for your entry" — build something like this, not rebuild this.

The single most useful fact in this table: Comfy's own Cinematic Studio reference app is prompt composition over a hosted model. It does not condition on camera, lens or light. The prior internal report found Higgsfield's own claim to condition on those panels is contested by a control experiment. Nobody in this competition has actually built the conditioning. That is the gap.

1.3 Rules, prizes, mechanics

The "$25,000" headline is a valuation; cash + credits total $16,500 plus three GPUs.

1.4 The licence trap — the thing almost nobody will check

"Open weights" ≠ "open licence".

Only the LTX cash bonus stacks cleanly with Open Weights Only. MiniMax H3 — which both comfy-reshoot and comfy-openjutsu depend on — is the trap: for a US entrant, the H3 bonus realistically means the hosted API, which forfeits Open Weights Only.

2. What we are building

2.1 The entry

An open-source ComfyUI app for re-shooting a clip from any angle, whose differentiator is a novel custom node pack that makes the camera move verifiable.

Working name: ComfyUI-Cinematographer — the missing DP layer for Comfy's own re-shoot app.

2.2 The node pack — the novel artefact

Nothing in the ComfyUI ecosystem evaluates a frame geometrically and tells you how many metres to move the camera. These nodes do:

The last one is the bridge. comfy-reshoot already shows a green/yellow/red reliability zone. Our node pack computes the framing quality on top of it — so the user sees both "is this inside the model's training range?" and "is this a good frame?", with a number for each.

2.3 The app

A director's panel where the user:

1. Loads a clip → depth is analysed once (MoGe-2)
2. Orbits / dollies / keyframes a camera move — the existing re-shoot UX
3. Presses Evaluate → the panel prints rule-by-rule arithmetic and correction levers
4. Presses Auto-Frame → the solver moves the camera to a compliant pose
5. Generates → LTX + CrossView-Warp renders the new angle
6. Checks continuity → Choros validates the 180° line and the 30° rule across the shot list

Step 3 is the thing no competitor has, and it is the thing that makes the demo watchable: numbers change on screen when the camera moves.

2.4 The model stack — licence-clean

Deliberately excluded: MiniMax H3 — US territory exclusion. This is a real differentiator against both reference apps.

3. What we already have — the "TON of ComfyUI functionality"

Full evidence in analysis/10-comfyui-substrate-inventory.md and analysis/11-contentstudio-inventory.md.

3.1 The ComfyUI integration — shipped

3.2 Verified workflow graphs — shipped and executed on real GPUs

services/local-resource-manager/comfyui_workflows/ — real ComfyUI API-format graphs, converted from ComfyUI's own templates against a live /object_info:

The _README blocks record provenance, exact template version, deliberate deltas from upstream, and — for the FLF graph — an explicit "deliberately NOT registered: the graph is verified, the request contract is not." That is the quality bar for the competition repo.

3.3 The creative engine — ContentStudio

internal/contentstudio/ — 440 Go files, ~131,000 lines, 47 http_*.go route files.

• The DLO (~3,680 LOC): metric scene graph in metres, real optics (50 mm lens, 36 mm sensor, 1.6 m eye height), USD layer stack, depth + object-ID passes, conditioned render
• The Cinematographer (~883 LOC): EvaluateFraming, AutoFrameCamera, Choros continuity
• A pure-Go software rasteriser with z-buffer and perspective-correct interpolation (aov_renderer.go:417-461) emitting metric depth, camera-space normals, instance IDs, mattes
• A 16-bit metric depth + object-ID pass with JSON sidecar (dlo_depth_pass.go)
• A node-graph engine (graph_run_*.go, http_graph_*.go) — flag-gated OFF by default (NODE39_GRAPH_RUNS, only ws-25), so the Run API 404s in production
• Poiesis (internal/poiesis/, ~12.6k LOC) — a separate workflow-graph engine with compile / compatibility / requirements / policy stages

3.4 The rulebook — the actual asset, verified

docs/features/cinematographer/cinematography-bible.md v1.3, 38 rules, each tagged DET (deterministic — computable from 3D coords + camera projection) or PER (perceptual — vision-model scored), each with a measurement/heuristic and a correction lever.

The prior report's four load-bearing claims, verified against the code:

This is the most important correction in the whole package. The honest number is 14 machine-evaluated rules, not 38 — and AutoFrameCamera returns an iterate, not a proven solution. Our competition entry must not repeat the overclaim. We will port the rules that are actually implemented, and we will add a convergence check to the solver. That is a strengthening, and it is exactly the kind of rigour that wins a judged competition.

3.5 The strategic prior art

../client-reports/higgsfield-cinema-studio-4.0/ is a complete adversarial teardown of the product we are asked to rebuild — including the finding that Cinema Studio's headline style-engine claim is contested by a control experiment, and that its three sponsored tutorials show two different builds. That report is our unfair advantage: we know exactly where the incumbent is weak, with receipts.

4. The mapping — our assets onto their cards

5. Why this wins — corrected for the competitive reality

5.1 The obvious ground is taken

So "we rebuilt Genjutsu" is not a winning pitch. Neither is "we rebuilt re-shoot" — Comfy already ships it. The differentiation has to be a novel node, a better interface, a missing mode, or an ignored category.

5.2 Our differentiation is a novel node — and it is the hardest visible problem

The re-shoot UX's hardest problem is where do I put the camera? The reference app answers it with a hard-coded green/yellow/red zone. We answer it with geometry: a measured pass/fail per rule, a correction lever in metres, and a solver that moves the camera for you.

That is:

• Novel — no framing-evaluation node exists in the ecosystem. CameraReference3D gives you 17 presets; we give you a judgement.
• Demonstrable in 30 seconds — move the camera, watch the numbers change. It is arithmetic, not a conditioning claim.
• A direct answer to the incumbent's weakness — Cinema Studio has no continuity enforcement at all, and Higgsfield's own consistency guide concedes the problem: "AI video distortion is not a quality problem. It is a memory problem… each new generation is a fresh context." Choros enforces the 180° line deterministically, across cuts.

5.3 The category fit

The honest read: we are not going to out-build Freefield's 95 tools in 15 hours. We are going to build one thing nobody has, on a licence-clean stack, and demonstrate it working. That is a category win, not a sweep.

5.4 The risk that dominates everything

Two search syntheses report that Higgsfield served Comfy a cease-and-desist on 10/28 evening (trademark + CEO likeness) and Comfy refused on 10/29. This is REPORTED, NOT INDEPENDENTLY VERIFIED — the primary Reddit disclosure is JS-blocked to our tooling. If true, the event may not proceed as advertised. We should not build anything that depends on Higgsfield's trademarks or likeness, and we should not name the entry after a Higgsfield product. Our working name — ComfyUI-Cinematographer — is deliberately clean.

6. What has to be built in the next ~15 hours

Everything above is Go, in a private monorepo, behind Magieva's auth, billing and org model. None of it is a ComfyUI node pack, and none of it is open-source today.

Critical path: 2 → 6 → 7 → 8 → 9. Items 1, 3, 4, 5 run in parallel with 2.

The defensible cut. 15 hours is not enough for a full port plus a polished app plus a deployment. The cut is: a genuinely working node pack (the 14 implemented rules + the solver + Choros + the reliability zone), one end-to-end workflow, and a panel that shows the arithmetic. That is a complete, honest, novel contribution. A half-ported pack with no working workflow is not.

What we will NOT do: claim 38 rules when 14 are implemented; claim the solver converges when it does not; depend on MiniMax H3; name anything after a Higgsfield product.

7. Decisions taken

8. The one thing still blocking

There is no ComfyUI login in the credentials file. It contains only the Magieva login (nathaniel@magieva.com + password). I need one of:

• the comfy.org credentials (if the account exists), or
• confirmation that nathaniel@magieva.com is the comfy.org account and I should sign in with the same password, or
• you handle the comfy.org signup + submission yourself, and I hand you a finished repo, video and submission text.

This does not block the build. It blocks only steps 8–9. I am proceeding with the build.

9. What happens next

1. Now — freeze the node-pack API in 01-SPEC.md
2. Then — parallel build: node pack · workflow graphs · panel app · repo scaffolding
3. Then — a verification lane that actually runs the thing end to end
4. Then — demo video + submission

Package contents