Submission and demo

03 — Submission checklist

Deadline: 2026-10-30 09:00 PT. Winners announced 11/6. Event: https://comfy.org/open-source-higgsfield/

The submission channel — verified

The form is https://form.typeform.com/to/V1dSQrkS. It is a Typeform, it needs no login, and it is the only submission channel. Verified live: isFormClosed: false, created 10/02, published 10/05, last updated 2026-10-26T17:46:34Z.

There is a separate form for event signup / free credits — https://form.typeform.com/to/Ayn9Cwfj. That is not the submission.

This retires the comfy.org-login blocker for filing purposes. The login was only ever needed to read the rules page. The form itself is anonymous. We can file the entry without a comfy.org account. A login is still useful for the Cloud validation run (see PLAN.md §4b), but it does not gate the submission.

Deadline: 2026-10-30 09:00 PT = 16:00 UTC = 12:00 ET. Hard Typeform close, no grace period. Winners announced 2026-11-06.

The 12 form fields, verbatim

(The form has 12 top-level questions; several expand into the sub-fields listed above.)

Field 7 — the one-line description, ≤100 characters

Camera moves you can verify: measured pass/fail per framing rule, with the correction in metres.

That is 99 characters. The max_length: 100 is enforced by the form, so do not paste a longer line.

Field 15 — categories, exact spelling

The form's multi-select offers exactly:

• Most Innovative
• Most Viral
• Open Weights Only
• Built with FLUX 3 Bonus
• VFX Built with LTX Bonus
• Built with MiniMax H3 Bonus

We select: Open Weights Only + VFX Built with LTX Bonus + Most Innovative.

Best Overall is NOT selectable. Every entry is auto-considered for it. Any earlier note that said otherwise was wrong. Bonus tracks stack with a main category.

The LTX bonus is $2,000 cash and requires that LTX be the core video model in the workflow — which it is in app/workflows/reshoot_ltx_crossview.json. Same form, no separate submission.

Field 17 — the four mandatory agreements

min=4 max=4, so all four must be ticked:

1. The repo is public and carries an open-source licence.
2. The work is lawful and SFW, with no unlicensed IP and no real person's likeness used without permission.
3. Comfy, Runpod, LTX, BFL and MiniMax may feature the submission per the Creator Feature Terms.
4. "This is my only submission."

Gap worth knowing: agreement 3 references "Creator Feature Terms" with no hyperlink, and no such page exists on comfy.org (404 on every guess; absent from the sitemap). We are agreeing to a document we cannot read. That is a real, if small, risk — recorded here rather than glossed.

What is NOT required

Correcting assumptions that would have cost us time:

• ❌ A ComfyUI custom node pack is not required.
• ❌ Comfy Registry publication is not required.
• ❌ ComfyUI Manager installability is not required.
• ❌ The demo video is a URL field only — no platform requirement, no public/unlisted rule, and the 2-minute limit is not mechanically enforced.
• ❌ No OSI-approval requirement is stated — "an open-source license" is the whole bar.
• ❌ No age, geographic, or employment restriction is stated. Country is collected but nothing is said about eligibility.
• ❌ There is no team field (a single Name and Email), so this is an individual entry.

What IS required beyond the form: the project must be built on the Comfy Developer Platform (Comfy API / Router / SDK). That is why field 12 matters and why the Cloud validation run in PLAN.md §4b is not optional polish — it is the evidence for that requirement.

Prizes — the real numbers

Total cash is $12,000. The "$25,000" headline is a valuation of the whole pool, not cash. No judges have been named and no rubric has been published. The only stated philosophy is: "We're judging the work, not your video editing or quality."

✅ The publication decision — DECIDED and EXECUTED

Repo: github.com/nathaniel-magieva/ComfyUI-Cinematographer (personal account, not the MagievaInc org). Public, Apache-2.0. The user chose the personal account explicitly.

The split is by branch, not by omission:

The user approved publishing the buildlog with its evidence layer, which supersedes the earlier "publish app/ only" recommendation. PLAN.md, 01-SPEC.md, 02-DEMO-SCRIPT.md, 03-SUBMISSION.md, 00-BREAKDOWN.md, verification/ and analysis/ ship on gh-pages.

verification/go-reference-dump.json IS published — it was scanned and is clean: 0 hits for internal/, rulebook, magieva-core, forge.magieva, MagievaInc, contentstudio, AI-DP. Its 84 non-ASCII strings are all rule target text, not internal paths.

What stays private: the Magieva monorepo itself, and any internal Go source path. The app/ tree has no Magieva imports and its tests run with zero Magieva dependency — a deliberate design constraint that pays off exactly here.

Scrub state (verified): grep -rniE 'internal/|rulebook|AI-DP|contentstudio|forge\.magieva|MagievaInc|magieva-core|[a-z_0-9]+\.go' app/ → no output. The only remaining Magieva occurrences are intentional brand/author: web/index.html (alt="Magieva"), pyproject.toml (authors, PublisherId).

The rules, as actually published

Finding: the competition page publishes no app requirements. All three accordions on comfy.org/open-source-higgsfield — "2. App requirements", "3. Prepare your submission", "4. Submit" — are empty in the served HTML. Verified by parsing the raw page: 3 accordion-content regions, all with empty bodies (<!--v-if-->). Only "1. Sign up" has content.

So the authoritative rule sources are the Typeform's four agreements, the blog post, and the rendered Q&A — not the page. We cannot be disqualified for breaking an unpublished rule, but we also cannot lean on one. The safe reading is the strictest published statement.

Is there a rule against improving the whole thing? No — it is the brief. From blog.comfy.org:

"Higgsfield promised to open source their entire platform. They didn't, so let's help them." … Build it. Publish the repo. Let anyone run it, fork it, and make it better.

The inspiration list on the page is the whole product, not one feature: "Cinematic Studio — Direct your shot like a film set / Re-shoot a video / Re-shoot from any angle / Openjutsu — Swap a character in a video." Rebuilding the full Cinema Studio flow is the brief.

Three constraints that do bind:

Constraint 2 is the one to watch. Our repo is a node pack + six graphs + a panel. That reads as a workflow export with a viewer unless the panel is a real, usable app surface. The Director's Panel is what makes it an app — which is why the Director tab and the full 5-stage flow matter more than any single engine module.

Unverified — do not repeat as rules. Two claims circulate in secondary write-ups that I could not confirm from any primary source: that a pure workflow export is disqualified, and that entries must introduce novel functionality or a distinct UI. The page's own "App requirements" section is empty, so neither is published. Treat them as likely intent, not as rules.

On the two competitor repos (sintetizacom/easycomfee, rossaai/rosesfield — both MIT, and rosesfield is explicitly an entry to this same competition): reading them, citing them, and reimplementing patterns is fine and smart. Absorbing their code or their entry is not — it collides with constraint 1, loses Most Innovative outright ("an app that's never been done"), and is trivially detectable. What we take is patterns, with attribution.

Draft social post

Every re-shoot tool asks where you want the camera. Ours tells you whether the frame is any good — and how to fix it.

ComfyUI-Cinematographer: 14 cinematography rules, each returning a measured value and a correction lever in metres. Plus a solver that reports whether it actually converged, and a continuity engine that catches the 180° line across cuts.

LTX 2.3 + CrossView-Warp + MoGe-2. No MiniMax H3 — the US is an excluded territory and local commercial use needs $5k/month. Every model here is one you can actually ship.

Apache-2.0. Repo in the reply. #OpenSourceHiggsfield

Pre-submission verification gate

Run these. Both must exit 0. If either fails, do not submit.

cd app && /tmp/cine-venv/bin/python -m pytest tests/ -q
# expect: 0 failed. (609 passed at last run -- the number grows as nodes land.
#         The gate is "0 failed", not a pinned count.)

/tmp/cine-venv/bin/python workflows/validate_graphs.py
# expect: PASS on every graph, exit 0. A NOTE naming declared-not-read
#         class_types is acceptable; a FAIL is not.

cd .. && /tmp/cine-venv/bin/python verification/differential_check.py verification/go-reference-dump.json
# expect: differential checks run: 289
#         ALL FIXTURES AGREE with the original Go implementation.

/tmp/cine-venv/bin/python app/web/parity_check.py
# expect: RESULT: PASS

Then confirm, by hand:

• [ ] The repo is public and the licence file is present.
• [ ] The README install steps work from a clean clone.
• [ ] The demo video is under 2:00.
• [ ] The social post contains the literal string #OpenSourceHiggsfield.
• [ ] No file name or user-facing string contains a Higgsfield trademark.
• [ ] No claim of "38 rules" anywhere. It is 14.
• [ ] No claim that the solver always converges.
• [ ] Every workflow graph's _README node states whether it has been run.
• [ ] No claim that the Director's Panel generates anything — it composes a plan.
• [ ] No claim that POST extends or edits the video — those two are plan-only by design.
• [ ] No claim that the judge reads rendered pixels — it judges the camera that was directed.

Honesty ledger — claims we may and may not make