The analysis

Open Source Higgsfield — competition brief

Prepared: 2026-10-29, ~16:15 America/Los_Angeles · Deadline: Friday 2026-10-30, 09:00 PT Countdown at time of writing: ~16 hours 45 minutes.

Headline summary

Comfy Org is running a two-week, $25,000-prize open-source build competition called Open Source Higgsfield. The premise, in Comfy's own words, is that "Higgsfield promised to open source their entire platform. They didn't, so let's help them." Entrants are asked to rebuild Higgsfield's most popular consumer apps — named explicitly as Genjutsu and Cinema Studio — as working, open-source apps on the Comfy Developer Platform, and to publish the repo.

Six things matter most for Magieva, and four of them contradict or complicate the event page's own summary:

1. The premise is a fact-check failure. Higgsfield never open-sourced its platform, and no source located shows it ever promised to. The September 2026 viral claim conflated Higgsfield's Apache-2.0 LLM-training/GPU-orchestration repo (higgsfield-ai/higgsfield, now archived) with its commercial studio, which remains closed. Comfy's own launch post states the non-open-sourcing as the reason for the competition. The competition is not affiliated with, endorsed by, or licensed by Higgsfield — the LA Build Night page says so explicitly.
2. There is an active legal dispute. Multiple secondary sources report that Higgsfield served Comfy Org with a cease-and-desist on the evening of 2026-10-28, demanding the challenge be taken down within 48 hours over trademark use and the likeness of CEO Alex Mashrabov, and that Comfy publicly refused on 2026-10-29 (today). I could not read the primary Reddit disclosure — it is JS-blocked to this tooling — so treat the C&D as REPORTED, NOT INDEPENDENTLY VERIFIED. It is nonetheless the single largest risk to the event proceeding as advertised.
3. The "four inspiration prompts" are three apps, and they are Comfy's own, not Higgsfield's. The page shows three cards — Cinematic Studio, Re-shoot a video, Openjutsu — each with a title and a tagline. The task's fourth string ("Re-shoot from any angle") is the Re-shoot card's tagline counted separately. All three are Comfy-built reference apps on comfy.org/hub, with public MIT repos under Comfy-Org/. They are the intended starting templates, not Higgsfield products. There is no Higgsfield app called "Re-shoot" — that name belongs to an unrelated third-party iOS app.
4. "Openjutsu" is a pun on Genjutsu. Higgsfield's Genjutsu (launched 2026-08-31) is a video-to-video character/object replacement app. Comfy's Openjutsu is its open rebuild. The LA Build Night page calls the app "Genjutsu"; the SF page and the hub call it "Openjutsu" — the same target under two names.
5. The bonus categories and the Open Weights Only category are in tension — and the licence axis makes it worse. FLUX 3 video is Cloud/API-only with "open weights coming soon", so a "Built with FLUX" entry cannot also be Open-Weights-Only. MiniMax H3 looks dual (open weights and a partner node) but its community licence excludes the United States as a territory and requires a $5,000/month licence for local commercial use — so the H3 bonus also points at the hosted path. Only the LTX bonus ($2,000 cash) stacks cleanly with Open Weights Only. Separately, the only fully unencumbered open video family is Wan 2.1/2.2 (Apache-2.0); LTX is revenue-capped at $10M and FLUX 3 Image Klein at ~$5M.
6. Higgsfield's actual model is a closed third-party one. Cinema Studio 4.0 runs on Seedance 2.5 (ByteDance) — stated in Higgsfield's own help centre. Every Higgsfield-branded model (Soul, Soul Cinema, DoP, Speak, Genjutsu, MS Image) is closed and proprietary, and Higgsfield has a real self-serve API (launched 2026-09-16) that resells Kling, Seedance, Wan 3.0, MiniMax H3 and LTX 2.5 Pro. There is no Higgsfield model or inference stack to fork — a rebuild is a from-scratch reimplementation of the product surface on open models.

Competitive reality check: at least one substantial rival entry is already public. Freefield (MIT) reportedly rebuilds ~95 of Higgsfield's ~129 tools as ~117–120 open-weight ComfyUI workflows, published from 2026-10-27, and a separate community repo D3voz/openjutsu implements the Genjutsu-style swap. Comfy's own comfy-openjutsu reference repo already has ~53 GitHub stars. The "obvious" rebuilds are taken; differentiation has to come from execution quality, the interface, or a category the crowd has ignored.

Rules and requirements

Sourced from the event page's rendered accordion content, which is not visible in the page's static text — it lives in Comfy's i18n bundle (eventChallenge.* keys in /_website/translations.BotHOeuL.js). The submission form's own field definitions corroborate it.

What the submission form actually asks for

Field-by-field, from the live Typeform definition (form.typeform.com/to/V1dSQrkS):

• Tell us about yourself — Name (required, "how you will be featured"), Comfy app Email (required, "results and prize shipping"), Country (required).
• Usernames — Social handles, Discord, Reddit (all optional).
• ComfyUI familiarity — required single choice: Never used it / Tried a few times / Used it regularly / Use it professionally / I build custom nodes.
• Project — Project name/title (required), one-line description ≤100 chars (required).
• Submission links — GitHub repo URL (required), Demo video URL (required, <2 min), Live link (optional), Social post link tagging #OpenSourceHiggsfield (required).
• Which parts of ComfyUI did you use? — required multi-select: Comfy API / Comfy Router / Comfy SDK / Comfy Cloud Agent / Comfy MCP.
• List all generative models used — required free text.
• Other tools or techniques — optional; "especially for the Most Innovative category".
• Category fit — required multi-select: Most Innovative / Most Viral / Open Weights Only / Built with FLUX 3 Bonus / VFX Built with LTX Bonus / Built with MiniMax H3 Bonus. (Best Overall is absent because it is automatic.)
• Is your stack fully open? — required: Yes / No / Not sure.
• Agreements — required multi-select of four checkboxes (public repo + OSI license; lawful/SFW/no unlicensed IP or likenesses; feature-rights grant; "This is my only submission").
• Anything else — optional.

Form defect worth noting: the Social post link field's helper text is a verbatim copy of the Live link field's text ("We know a live link costs you money… Expect roughly 5-10 runs…"). The social-post field has no correct description. Cosmetic, but it is evidence the form was assembled quickly.

What "sign up" gives you

• The signup form (form.typeform.com/to/Ayn9Cwfj) asks: name, "Your Comfy account email — This is where the credits land, so it has to be the real one", ComfyUI depth, Build Night attendance, and an optional one-liner on what you'll build.
• Thank-you screen: "Credits are on the way if you made the first 200." Welcome screen: "Free credits for the first 200. Show up with nothing but an idea."
• The amount is not published anywhere. I checked the event page, both forms, the launch blog post, both Luma pages, and searched for it. The only credit figures Comfy states are the runner-up prize ($250 Comfy Credits) and the partner bonuses. Treat "how many credits" as UNKNOWN. The signup form's own thank-you also promises "We will remind you two days before" the deadline — two days before 10/30 was 10/28, so that reminder window has already passed.
• Where they are usable: credits land on the Comfy account (the email is the delivery address). Comfy Cloud credits and Partner Node credits are the same pool — "API requests draw from the same monthly credit allocation as the Comfy Cloud web UI. There is no separate API credit pool." Partner Nodes are also billable from a local ComfyUI install if you supply an API key in the prompt payload, so the credits are plausibly usable for partner-node calls from local ComfyUI too. UNVERIFIED whether the signup grant is restricted to Cloud GPU time, partner nodes, or both.

Prizes and categories

Total advertised: $25,000 ("total worth in prizes"). Note the arithmetic: $10,000 + $2,000 + $1,000 + $1,000 + (10 × $250 = $2,500) = $16,500 cash/credits, plus 3× RTX 5090 and the FLUX tiering. The $25,000 headline is a valuation, not a cash pool.

Stacking: "Each entry is eligible for both a main category and a model-provider bonus category." So one entry can win Best Overall and an LTX bonus. It is not stated whether one entry can win multiple main categories (e.g. Most Innovative and Open Weights Only) — assume not, but UNVERIFIED.

Contradictions and gaps in the prize copy:

• The event page says "Flux API credits"; the launch blog says "FLUX API credits"; the submission form's category label says "Built with FLUX 3 Bonus". Same thing, three spellings.
• The LA Luma page says: "$10,000 Grand Prize + 3 GeForce RTX 5090s + bonus prizes from OSS ecosystem partners LTX and BFL" — MiniMax is omitted there, though the SF Luma page and the event page include it. The LA page also adds "special in-person awards" for in-person entries, which appear nowhere on the official rules page. These in-person awards are undefined and unverifiable.
• The launch blog says the competition is "Backed by Runpod, LTX, Black Forest Labs, and MiniMax"; the wiki adds NVIDIA; the event page footer's rights grant names ComfyUI, Runpod, LTX, BFL, MiniMax (no NVIDIA). Partner list is inconsistent across sources.
• The blog headline says "Build with us IRL in SF, NYC & LA" but only two Build Nights exist (SF 10/28, LA 10/29). No NYC event was found. The blog also says "We're hosting 2 Build Nights" in the body — the headline is wrong.
• The event page's own "4 inspiration prompts" framing is wrong — see below.

Judging

Who judges: the internal Comfy Org team. No named judges, no published panel, no scoring rubric, no weighting, and no tie-break rule were found on any Comfy page, form, or the launch post. Multiple secondary sources describe it as Comfy-team judging; none name an individual. Treat the judging panel as UNPUBLISHED.

The only published judging philosophy is the FAQ answer on demo videos:

"We're judging the work, not your video editing or quality. The best demo shows your work in action with an explanation of how it was built."

What that implies operationally (inference, not stated):

• A working app beats a beautiful video. But the video is the only artefact guaranteed to be watched, so it must show the app running end-to-end.
• The live link is the strongest evidence and the judges will use it 5–10 times. A flaky or credit-exhausted live link is worse than none.
• "Most Innovative" is the only category with a stated evidentiary requirement (the video must argue the innovation).
• "Open Weights Only" is the only category with a stated verifiable test (the stack-openness question on the form). It is the most objectively winnable category and the easiest to prove.

The four inspiration prompts, decoded

Correction to the brief's framing. The event page does not present four prompts. It presents three app cards, each with a title and a tagline:

The task's list of four splits card 2 into its title and tagline. There is no fourth app. The section's own lead text describes the space, not four products: "Start with a recognizable output: a performance effect, a cinematic camera move, or a music-video transformation."

Critically, all three are Comfy's own reference apps, not Higgsfield products. Each has a public MIT repo under Comfy-Org/, described as "clonable … and development guide". They are the intended templates. The hub listing labels them "Comfy app". The event page's own copy calls them "inspiration prompts for your entry" — i.e. build something like this, not rebuild this Higgsfield product.

Cinematic Studio — "Direct your shot like a film set"

• What it is: Comfy's own hub app (Beta). Hub summary: "Direct a shot with a real camera, light and grade, then run it on any image model." Metadata: Image · 5 models. App description: "Direct cinematic stills: pick the camera, shot, light, film and grade, then run any image model through the Comfy Router."
• Reference repo: Comfy-Org/comfy-cinematic-studio — MIT, ~10 stars. "A clonable Comfy Router app and development guide."
• How it actually works (from the repo README): a Next.js app. It does not submit a ComfyUI workflow at all. app/api/generate/route.ts calls bfl/flux-2-pro through the Comfy SDK's Router surface. lib/prompt.ts's cinematicPrompt() composes the scene text plus the selected camera/light/look into one prompt string. The response parser reads an HTTPS image URL from result.sample. So the "director's panel" is prompt engineering, not a conditioning model. That is a materially different architecture from Higgsfield's Cinema Studio, which claims the parameters are locked as generation parameters.
• The control surface (from Comfy's i18n bundle): camera body (Digital cinema, Large format, Super 35, 35mm film, 16mm film, Handheld), lens (Spherical, Anamorphic, Vintage, Macro, Tilt-shift), focal length (14/24/35/50/85/135 mm), aperture (f/1.4, f/2, f/2.8, f/4, f/8), framing (Extreme wide, Wide, Medium, Close-up, Extreme close-up, Over the shoulder, Low angle), light (Golden hour, Overcast, Blue hour, Practical night, Neon, Low key, Silhouette), film stock (Digital clean, Tungsten 500T, Daylight 250D, B&W 400, Reversal slide, Expired film, Bleach bypass), look/genre (Neo-noir, Western, Science fiction, Period drama, Thriller, Documentary, Road movie), grade (Teal and orange, Noir, Kodachrome, Neon night, Desert dust, Nordic cold). Aspect: Scope, Widescreen, Academy, Landscape photo, Portrait photo, Square, Vertical. There is also a video mode (creation type image/video/animate image, duration, audio, first/last frame, source video) and a colour-palette tool ("Match the colors of an image", hue/saturation/hex editing).
• Rebuild difficulty: LOW. This is a prompt-composition app over one hosted model. A credible clone is a weekend. The hard part is not the app — it is that the category is crowded and the "director's panel conditions the output" claim is unproven (see the internal report).

Re-shoot a video — "Re-shoot from any angle"

• What it is: Comfy's own hub app, marked Prototype. Hub summary: "Aim a new camera at your clip and generate the scene from that angle." Metadata: Video · MiniMax H3.
• Important: there is no Higgsfield app called "Re-shoot." That name belongs to an unrelated third-party iOS app (ReShoot: AI Face Photo Editor by HUBX). This card is Comfy's own app name, not a Higgsfield product. The closest Higgsfield features are Angles 2.0 (12 auto-views from one image), Shots (9 shots from one image) and Zooms — all image-side — while the video-side equivalent is Genjutsu's motion transfer. So Comfy's Re-shoot is closer to a new capability than a rebuild of a named Higgsfield app, which is part of why it is the most technically interesting of the three cards.
• Reference repo: Comfy-Org/comfy-reshoot — MIT, ~6 stars. "A clonable Comfy workflow deployment app and development guide."
• How it actually works (from the repo README): a Next.js app over a Comfy Developer Platform deployment (Comfy API v2 via the SDK). Two workflows:
• workflows/analyze.api.json → "Analyze depth", returns a .cvgeo depth file for the browser preview.
• workflows/generate.api.json → "Generate new view", which runs MoGe + CrossView Warp + MiniMax H3 and saves three videos (generated result, an original-audio version, and a warp guide).
• lib/comfy.ts maps the uploaded clip → node 1, output settings → node 2, camera values → node 5, prompt → node 20.
• Deployment needs the custom CrossView nodes, MiniMax H3, MoGe, and the CrossView LoRA. Demonstrated on a 96 GB GPU; smaller configs untested. Sample clip: samples/sunlit-room-test.mp4 (5 s). Limits: MP4 ≤100 MB, clips 5–15 s, changing clip/aspect/size forces re-analysis.
• The app's own credit line, from Comfy's i18n bundle: "Built on Cseti's CrossView-Warp LoRA and node." This is the load-bearing technical fact.
• CrossView-Warp (verified independently): an IC-LoRA for LTX-Video 2.3 (22B) by creator Cseti (cseti007). HF model Cseti/LTX2.3-22B_IC-LoRA-CrossView-Warp_v2 (v2 2026-08-15; v0.9 2026-07-28). ComfyUI node pack: github.com/cseti007/ComfyUI-CrossViewWarp. It estimates depth with MoGe-2 (or Depth Anything V2), reprojects frames into 3D from the requested camera pose, marks disoccluded regions magenta, and feeds two in-context guides to LTX — the depth-warped clip (camera/geometry) and the original clip (identity/texture/lighting). Trigger word crossview; latent_downscale_factor = 1 on both guides; LoRA strength ~1.0–1.15 rigid / 1.2–1.3 human.
• Note the model mismatch: the hub app is tagged MiniMax H3, but its credit line and its dependency list are LTX 2.3 + CrossView-Warp. The most likely reading is that CrossView-Warp produces the warp guide and MiniMax H3 performs the actual generation (H3 is the app's video model), with MoGe for depth. UNVERIFIED which model does the final pixels.
• The UI (from i18n): two steps — "Prepare the clip" (choose clip, framing, what the new view reveals; depth analyzed once) then "Aim and generate" (drag to orbit, scroll to dolly, keyframe a move, generate). Controls: rotation, tilt, lens, height, azimuth, elevation, distance, FOV, vertical shift. Views: Result / Warp guide / Source. Sound: Generated or Original audio. A reliability zone indicator: green = "Inside the range the LoRA was checked on", yellow = "Trained, but less reliable", red = "Outside training. The hidden side will be invented." Keyframing supports linear / ease-in / ease-out / ease-in-out / smooth-spline. Output 480p (0.4 MP, 864×480, faster) or 768p (1.0 MP, 1376×768, sharper/slower), 24 fps, 5–15 s. Depth analysis takes 20–40 s.
• Rebuild difficulty: MEDIUM–HIGH. The reference app is clonable, but the hard dependencies are a 96 GB GPU, a custom node pack, a third-party IC-LoRA, and a depth model. The genuinely hard engineering is the interactive 3D orbit preview and the keyframe interpolation, not the graph.

Openjutsu — "Swap a character in a video"

• What it is: Comfy's own hub app. Hub summary: "Swap a character in your video for one from a single image, and keep the scene and its sound." Metadata: Video · MiniMax H3.
• Reference repo: Comfy-Org/comfy-openjutsu — MIT, ~53 stars (the most popular of the three). "A clonable Comfy character-swap app and development guide."
• The name: a pun on Higgsfield Genjutsu. The LA Build Night page names the target app "Genjutsu"; the SF page and the hub name Comfy's version "Openjutsu". A separate community repo, D3voz/openjutsu, describes itself as "One-click character swap for videos: an open-source, local Higgsfield Genjutsu alternative built on ComfyUI and MiniMax H3, with FLUX.3 Klein relighting."
• How it actually works (from the repo README): a Next.js app over a Comfy Developer Platform deployment.
• workflows/swap.api.json — clip → node 30, character image → node 9, part start/length → trim node and node 31, quality → node 31 (480 or 768), prompt + seed → node 801. Node 809 saves the result.
• Built on Akatz Labs' H3 Character Swap workflow (akatz-labs/h3-character-swap-tool at commit 0670e4f), with three documented changes: swap LoRA v1 → h3_character_swap_v1.5; add a Video Slice trim node; replace the anchor node with LoadImage + ImageScale.
• Deployment needs: a patched h3_character_swap node pack, MiniMax H3, SAM 3.1, and two LoRAs. ComfyUI 0.41.2+ (core provides MiniMax H3, SAM 3.1, TextGenerate). Run on a 96 GB GPU (RTX PRO 6000). A 5 s part at 480p took ~2 min on a warm worker.
• The prompt template: "Replace the selected person in <Video 1> with the character in <Picture 1>. Preserve the source performance, scene and audio timing." — H3 reference tags, plus optional extra direction.
• The optional first-frame "anchor": before the swap, the app posts the part's first frame plus the character image to /api/anchor, which asks GPT Image 2.5 (openai/gpt-image-2.5-sunburst, quality low) through Comfy Router to composite the character into that frame; the swap then starts from that image. This is a closed API in the critical path — so the reference app as shipped is not Open-Weights-Only eligible. If the anchor fails or is unticked, those nodes are removed.
• The UI (from i18n): upload video (MP4/MOV, up to 15 s used), add a character image, optional extra direction ("darker jacket, keep the glasses, less smiling"), seed, quality 768p/480p, "Swap character". "The original sound is kept on the result. Each run also makes a first-frame image, charged separately." Four built-in examples: singer/Home studio/Disco diva, boxer/Old gym/Comfy fan, painter/Painter's studio/Ronin, cowboy/Desert canyon/Robot.
• Rebuild difficulty: MEDIUM. The reference app is clonable and the workflow is a fork of a public one. The real work is the deployment (patched node pack + 96 GB GPU + two LoRAs) and the optional anchor step. Dropping the GPT Image 2.5 anchor is the single cheapest way to make this entry Open-Weights-Only eligible — and the repo already supports running without it.

Higgsfield apps to rebuild

Higgsfield's product surface, from higgsfield.ai/apps (48 app pages) plus its product pages. Marketing copy is labelled as such; model attributions are the weakest claims here.

Cinema Studio (4.0)

• What it does: Higgsfield's flagship cinematic studio — a pre-generation directorial control surface simulating a film set: camera/sensor profiles, lens optics, focal length, aperture, colour grading, physics-aware motion, native audio, an AI Director, and an "Elements" asset system. Higgsfield's own page: "Control the camera, lighting, color, and acting in every shot with 30+ camera movement presets and 50+ color palettes. Edit each scene independently and extend your story in any direction." (marketing)
• Versions: 2.0, 2.5, 3.0, 3.5, 4.0. 4.0 loads by default; a selector switches 2.0–4.0. 2.5 added up to 3 Soul Cast characters and built-in grading (colour temperature, contrast, saturation, sharpness, film grain, highlights, exposure). 3.0 added a reasoning engine, up to 9 references, 7 genres, speed ramps (Linear/Auto/Flash In/Flash Out/Slow-mo/Bullet Time/Impact/Ramp Up), native audio and a content-protection filter. 3.5 added the Claude-powered AI Director (adjusts genre/style/camera and breaks scripts into shots; never triggers generation itself), the Elements system, team collaboration, per-shot camera settings, 8 colour palettes, 10 Camera MoveSet styles and 6 lighting presets. 4.0 adds the Director's Panel, up to 50 references, and clips up to 30 seconds.
• Inputs: prompt/script; up to 50 references (4.0); Elements (characters/locations/props) referenced with @; Soul Cast characters; Director's Panel controls — Film Setup (genre, era 1960s–2020s, tempo Single Shot→Chaotic), Camera (35mm Film / 8mm Film / DV Camcorder, lens clean-sharp→anamorphic/vintage, aperture, moves including POV, Robot Arm, Helicopter Shot), 50+ colour-grade templates, 6 lighting presets.
• Output: video with native synchronised audio (SFX, speech and music in the same pass), up to 30 s, up to 1080p (4K on some underlying models).
• Underlying model: Cinema Studio 4.0 runs on Seedance 2.5 (ByteDance) — stated explicitly in Higgsfield's own help centre. Earlier versions are Higgsfield-branded with no published architecture. "Cinema Studio" also appears as its own line item in credit tables, so it is sold as a model-like unit. This is the single most important model attribution in this brief, and it is a closed third-party model.
• Pricing: conflicting published numbers. Higgsfield's own 2026-10-20 blog: 16 credits/sec at 1080p (160 per 10 s); 3 credits at 480p, 7 at 720p. A third-party teardown of the pricing page: 25 credits/5 s @720p, 50/5 s @1080p, 50/10 s @720p, 100/10 s @1080p. Treat the per-second figure as current.
• ComfyUI rebuild difficulty: LOW technically, HIGH to differentiate. Comfy's own comfy-cinematic-studio is prompt composition over bfl/flux-2-pro — i.e. it does not reproduce the claimed conditioning. The internal report documents a public API schema with closed enumerations (4 camera bodies, 5 lenses, 3 apertures, 33 camera moves, 6 genres, 5 eras, 4 pacings, 6 lighting presets, 50 colour palettes) and finds that the claim the panels condition the output is contradicted by one of three primary sources, with a control experiment, while the two sources that demonstrate it working run demos whose prompts already contain the variable under test. A rebuild that actually conditions on camera/lens/light — via a real conditioning model rather than prompt text — would be more faithful than Comfy's own reference app, and is the obvious innovation angle. Do not duplicate the internal report — see client-reports/higgsfield-cinema-studio-4.0/analysis/00-SYNTHESIS.md for the full evidence base.

Genjutsu / Genjutsu Lite — the app behind "Openjutsu"

• What it does: video-to-video targeted replacement. Higgsfield's launch guide (2026-08-31, marketing): "Genjutsu takes an existing video and rebuilds part of it… it can change the person in the shot, the setting, or an object in the frame. Everything else stays exactly as filmed." Two modes: Motion Transfer (apply a clip's motion to a new character/scene) and Object Swap (replace a specified element, preserve the rest). A Motion Library of trending dances, transitions and memes is included.
• Inputs: one video 4–30 s (or a Motion Library pick); up to 30 reference photos; prompt optional (30+ presets). Genjutsu Lite (2026-10-20): 4–30 s, up to 10 references, 720p/1080p, no mode choice, ~⅓ the credit cost.
• Output: up to 1080p; preserves camera movement, lighting, grain, lens character and cut points. 480p/720p/1080p.
• Pricing: Genjutsu Lite 15 s @720p = 32 credits. Reported ~40/104/144 credits for 15 s @480/720/1080p (UNVERIFIED). API: from $0.318/sec (discounted $0.159/sec).
• Underlying model: not publicly disclosed. Higgsfield routes across open-weight and closed models; Genjutsu is a proprietary video-to-video pipeline. Reddit speculation points at Seedance 2.5; Higgsfield has not confirmed. UNVERIFIED. The closest open equivalents are SCAIL-2 (Wan 2.1 14B, replacement mode) and Wan-Animate-2.
• ComfyUI rebuild difficulty: MEDIUM. Comfy's comfy-openjutsu already does it. The differentiators are (a) dropping the closed GPT Image 2.5 anchor to qualify for Open Weights Only, (b) supporting the Object Swap mode (Comfy's app is person-swap only), and (c) the 30-reference / preset / Motion Library surface, which Comfy's app does not have.

The camera-angle family — Angles 2.0, Shots, Zooms, Camera Motion

Important correction: there is no Higgsfield app named "Re-shoot." That name belongs to an unrelated third-party iOS app (ReShoot: AI Face Photo Editor by HUBX). The competition's "Re-shoot a video" card is Comfy's own app name, not a Higgsfield product. The real Higgsfield equivalents are:

• Angles / Angles 2.0 (/apps/angles, under Edit Image) — "Generate new camera angles from a single photo." Full 360° control: front/side/back/overhead, "Generate from all angles" produces 12 auto-views in one click, Quick Select presets, an interactive 3D rotation disc, manual sliders (rotation, tilt/vertical, zoom), brightness and colour lighting control, and on-page generation history. Angles V2 shipped Nov 2025. Image, not video.
• Shots (/apps/shots) — "9 unique shots from one image." Image/storyboard. Cinema Studio 2.0 has an equivalent 9-shot 3×3 contact-sheet multi-angle mode.
• Zooms (Pro) — 9 zoom levels from one image.
• Camera Motion (/apps/camera-motion) — "Generate shots, angles, and seamless transitions."
• Underlying model: not published; Higgsfield-branded. UNVERIFIED. Angles-type products are typically image-diffusion plus camera-pose conditioning (a Qwen-Image-Edit or FLUX Kontext-style edit model with pose control).
• Rebuild difficulty: LOW–MEDIUM. ComfyUI has native camera-pose nodes (WanCameraEmbedding, WanCameraImageToVideo) for the video case; the still-image case is an edit-model + prompt/pose problem. Note ComfyUI_CameraAngleSelector already does 96 predefined angles for a Qwen-Image-Edit LoRA, so the image-side is largely solved.

Recast / Character Swap 2.0 / Face Swap — three distinct tools

• AI Face Swap (/apps/face-swap, /apps/video-face-swap) — replaces only the face. Image mode and video mode; video mode preserves original motion, lip sync and performance. No prompt, no mask, no retouch. Higgsfield's marketing claims "generative identity embedding", lighting/shadow matching and facial-geometry alignment — marketing, not verified.
• Character Swap / Instadump (/apps/character-swap) — transfers the full character (body, outfit, skin tone, persona, emotion), not just the face. 30+ ready-made aesthetic/CGI presets.
• Recast (/apps/recast) — character swap in video, plus instant voice cloning (upload a quiet-room audio excerpt), dubbing, translation, gender change, and background replacement (Cozy Minimalist, Pink Room, Luxury Masculine, Green Screen). 30+ character presets.
• Inputs: base/target image or video + one clear front-facing reference portrait (or a full-character reference / preset). Recast adds voice settings. Output: swapped image or video; video keeps original motion and audio unless dubbed.
• Underlying model: not published. UNVERIFIED. A Higgsfield-proprietary pipeline over third-party video models.
• Rebuild difficulty: LOW (face swap) to MEDIUM (full performance swap). Face swap is a solved ComfyUI problem (ReActor, PuLID, InstantID); performance swap is the Genjutsu/SCAIL-2 problem.

Soul — the image model family

• What it does: Higgsfield's proprietary image foundation family, sold as a model and as a studio. "Higgsfield Soul – Hyper-Realistic, Fashion-Grade AI Photo Model… 50+ aesthetic presets." (marketing) Components:
• Soul — quick preset-driven generation. Preset categories: General, TikTok Core, Instagram Aesthetics, Beauty, Mood, Camera Photo, Graphic Art. Avatar support.
• Soul 2.0 — full control: presets, Moodboards, Soul HEX, Soul ID, reference image. 33 curated presets. Quality 1.5K/2K, batch 1–4.
• Soul Cinema — cinematic film stills: film-stock grain, motivated practical lighting, cinema colour grades, lens depth. Soul HEX + Soul ID; no presets or Moodboards.
• Soul ID — a trained identity layer. Upload 20+ photos, trains in ~3–5 min, reusable across the whole Soul family. Solves "face drift".
• Soul HEX — colour transfer: extract a palette from a reference image, or use built-ins (Film colors, Lime Jam, Candy pink, Nostalgic blue, Soft palette, Black gloss).
• Moodboards — 20+ reference images (ideally faceless) → a reusable style preset. Soul 2.0 only.
• Also referenced (third-party, UNVERIFIED on higgsfield.ai): Soul Cast (character casting sheet: front/back/close-up) and Soul Location (environments, no people).
• Inputs: text prompt and/or reference image — mutually exclusive (attaching a reference disables the prompt field); preset or Moodboard; Soul ID character; Soul HEX palette. Aspect ratios 9:16, 3:4, 2:3, 1:1, 4:3, 16:9, 3:2. Output: still images, up to 4 per generation, 1.5K or 2K.
• Underlying model: Higgsfield-proprietary. Soul 2 / Soul Cinema are the API model names. One third-party review claims Soul Cinema uses "Nana Banana Pro" — conflicting and UNVERIFIED; Higgsfield's own docs present Soul Cinema as its own model.
• Pricing: Soul 0.25 credits/image; Soul 2.0 and Soul Cinema 0.12 credits/image (third-party teardown). Higgsfield's own blog: Soul 2.0 @2K = 0.125 credits (≈$0.006). API: Soul 2 and Soul Cinema $0.0032/image.
• Rebuild difficulty: MEDIUM. The app is easy; matching the aesthetic is a LoRA/preset-curation problem, not an architecture problem. Soul ID is the interesting part — a reusable trained identity is exactly what ComfyUI's IC-LoRA / ID-LoRA / PuLID route provides.

Speak — talking avatar / lip sync

• What it does: the lip-sync engine inside Lipsync Studio (Video → Lipsync Studio, "Create Talking Clips"). The standalone /speak page is only a gallery of themed avatar collections (Podcast, Coaching, Reporter). Higgsfield Speak 2.0 is the in-house model; the studio also offers Google Veo 3.1, Kling 2.6 Lipsync, Wan 2.5 Speak, Kling Avatars 2.0, Infinite Talk, Kling Lipsync (v2v) and Sync Lipsync 3 (v2v, up to 4K). Scene templates: General, Selfie, Podcast, Car Talking.
• Inputs: character image (front-facing, up to 2K, closed mouth, no obstructions) or video; audio — either TTS from an "Audio text" field (with emotion brackets like [excited]) or an uploaded MP3/WAV; a scene prompt for background/movement/camera only (dialogue must not go here); quality/resolution. Output: talking-video MP4, up to 1080p. A separate Audio workspace handles Text to Speech, Voice Change, Translate (18 languages) and custom voice cloning.
• Underlying model: Higgsfield Speak 2.0 — proprietary, architecture not published. UNVERIFIED. One independent reviewer observed a "Higgsfield DoP" label on the General template card.
• Pricing: Higgsfield Speak 2.0 ≈14 credits / 5 s @720p. Wan 2.5 Speak ≈6 credits / 5 s @480p, ≈20 / 5 s @1080p. Speak requires Plus or above (Free and Starter are marked ✗).
• Rebuild difficulty: MEDIUM. ComfyUI has the pieces (Wan 2.2 S2V, LatentSync-class, InfiniteTalk-class nodes). Not researched in depth here.

Popcorn — storyboard / multi-shot

• What it does: a native AI storyboard generator producing coherent sequences of up to 8 frames where character identity, lighting, atmosphere, clothing, depth, reflections and spatial logic persist across shots. Can mix real photos with AI images. Three modes: Manual (direct each frame), Auto (one prompt → up to 8 frames), Instadump (upload photos → styled photo set; a separate 3-step flow, packs built from up to 15 reference photos).
• Inputs: text prompt; up to 4 image references (addressed by number in the prompt); mode; frame count; aspect ratio (3:4, 2:3, 3:2, 4:3, 16:9, 1:1, 9:16, or Auto). Output: 1–8 connected storyboard frames (images). Frames can then be animated with Kling or Veo as start/reference frames; chaining the last frame of one clip into the next keeps motion continuous.
• Underlying model: Higgsfield-native; not published. UNVERIFIED.
• Pricing: 1.5 credits per Popcorn image (third-party teardown). Derived: 4-frame ≈ 6 credits, 8-frame ≈ 12 credits.
• Rebuild difficulty: MEDIUM. The interesting engineering is cross-frame identity and lighting consistency — which in ComfyUI terms is reference conditioning (FLUX 3 Image Klein multi-reference, or an IC-LoRA) plus a consistent-seed/latent-chaining strategy.

Effects — the VFX preset library

• What it does: "Explore cinematic AI visual effects from Higgsfield. Choose an effect, add your media, and create polished, share-ready videos in seconds." (marketing) 60+ ready-made one-tap "Viral Presets." Pick preset → upload photo → generate. No prompt, no camera setup. Each preset bundles animation, location and style. A separate Camera Controls library has 50+ named moves.
• Named presets (from higgsfield.ai/effects): Floating fall, High flip, Burning man, Scrapbook collage, Comic, Frozen in motion, Particles, Pigeons, Pearl earring, Studio slide, Incline, Act natural, Monster dab, Clones, Earth zoom, Cutout, Mighty fighter, Vanish, Moonwalk, Eyes in, Street colossus, Melting, Architecture wave, Agamemnon, Cyclope, Canvas, Tracking, LSD, Blue depth, Wild ride, World morphing, Stop world, Bullet time, Infinite clones, Superstar, Skatedog, Smash and grab, Selfception, Lacewalker, Boarding pass, Lidar transition, Fallen angel, Windows, Fairytale castle, Palette.
• Camera Controls (from higgsfield.ai/camera-controls) — the most useful enumeration in this brief for a camera-move rebuild: Static, Pan Left/Right, Tilt Up/Down, Dolly In/Out/Left/Right, Dolly Zoom In/Out, Super Dolly In/Out, Crash Zoom In/Out, Rapid Zoom In/Out, Crane Up/Down/Over The Head, Jib Up/Down, Arc Left/Right, 360 Orbit, 3D Rotation, Aerial Pullback, FPV Drone, Hyperlapse, Timelapse (Glam/Human/Landscape), Snorricam, Robo Arm, Whip Pan, Wiggle, YoYo, Fisheye, Dutch Angle, Low Shutter, Overhead, Object POV, Through Object In/Out, Focus Change, Head Tracking, Hero Cam, Road Rush, Buckle Up, Car Chasing, Car Grip, Lazy Susan, Handheld, BTS, Eating Zoom, Glam, Flying Cam Transition, Incline, Mouth In, Eyes In.
• Inputs: source photo (some presets accept video), preset choice, output resolution 480p–4K, aspect ratio (9:16 for TikTok/Reels/Shorts where offered). Output: short styled video clip.
• Underlying model: routes to underlying video models; not published per preset. UNVERIFIED.
• Rebuild difficulty: LOW per effect, HIGH to match catalogue breadth. This is where a LoRA library and a preset picker win. The Camera Controls list maps almost one-to-one onto ComfyUI-CameraReference3D's 17 base moves plus its timeline DSL — a strong, checkable rebuild target.

Draw-to-Video — exists, renamed

• Correction: this product does exist. The slug higgsfield.ai/draw-to-video returns 404, but the product was renamed. Current branding is Sketch to Video AI (/sketch-to-video-ai) and the app Sketch to Real (/apps/sketch-to-real). "Draw to Video" was the Aug 2025 name.
• What it does: you annotate directly on a canvas to direct motion — arrows (direction + speed, longer = stronger), curves (camera/orbital paths), small circles/loops (idle cycles: hair, cloth, breathing), numbered steps (sequence), bounding boxes plus text notes for per-object actions.
• Inputs: sketch/drawing/storyboard/line art, or an uploaded base image; drawn motion paths; optional text prompt (style/vibe); optional motion preset (e.g. Earth Zoom Out); model choice. Output: video clip; 16:9, 9:16, 1:1. Known defect: the first frame renders with your annotations visible — creators trim frame 1.
• Underlying model: routes through Higgsfield's multi-model workspace — reported engines Veo 3/3.1, Hailuo/MiniMax, Seedance, Sora 2, Kling, Wan. No fixed published model. UNVERIFIED at the per-tool level.
• Rebuild difficulty: MEDIUM. The novel part is the canvas annotation → motion-vector translation, which is a genuine UI/engineering problem and a plausible "Most Innovative" angle.

Product Photoshoot — exists as Marketing Studio

• Correction: this product does exist, just not under that slug. It lives inside Marketing Studio (/marketing-studio-intro) and also ships as the agent skill higgsfield-product-photoshoot in the official higgsfield-ai/skills repo.
• What it does: Marketing Studio has 1,500+ templates across six categories (Product shots, Motion, UGC, Ads, Posters, Marketplace). The core mechanic is "Recreate": open a template, swap in your product, pick an avatar, describe what should change — the template's layout/lighting/style regenerates around your product. Posters open as editable layers (headline, date, image each separate). Product-shot formats: Studio (seamless/marketplace white or transparent PNG), Lifestyle, With Model/On-Model, Closeup, Faceless, Full body, Editorial, UGC. Avatars: 100+ library, or generate your own via Soul 2.0 / Soul ID. The CLI skill exposes modes: product_shot, lifestyle_scene, closeup_product_with_person, moodboard_pin, hero_banner, social_carousel, ad_creative_pack, virtual_model_tryout, conceptual_product, restyle.
• Inputs: upload clean product photo(s) (multiple angles improve consistency) or paste a website/social product URL (auto-fetches name, description, images, logo, colours, copy); optional avatar; optional edit description; or a free-text prompt path. Output: product images up to 4K upscale; Marketing Studio also outputs UGC videos up to 15 s.
• Underlying model: Web UI → "MS Image" / Marketing Studio Image (Higgsfield-proprietary). CLI/agent skill → gpt_image_2 (GPT Image 2) — the backend prompt enhancer assembles the prompt and submits to gpt_image_2. API model name: Marketing Studio Image.
• Pricing: Marketing Studio UGC ad 10 s @720p = 70 credits (160 @1080p per one source). API: Marketing Studio Image $0.0059/image.
• Rebuild difficulty: LOW–MEDIUM. Template + reference-image compositing. The URL-ingestion path is the differentiating engineering.

Other named Higgsfield apps (competition targets)

Full nav as fetched from higgsfield.ai on 2026-10-29: Explore · Image · Video · Audio · MCP · API · Genjutsu (New) · Content Factory (New) · Playable (New) · AI Influencer · Academy (New) · Katana · ChatGPT Plugin · Ads Studio · Effects · Cinema Studio · Contests · Marketing Studio · Supercomputer · Edit · Community · Plugins · Canvas · Originals · Pricing · Enterprise. Studios: Cinema Studio, Marketing Studio, Lipsync Studio, Higgsfield Popcorn, Higgsfield Canvas, Soul, Soul 2.0, Soul ID, Character, Soul Cinema. Platform: Supercomputer, AI Motion Designer, MCP/CLI, API, Collab, Games, Reference, Extension.

Also in the nav but not detailed: Inpaint, Upscale, Sora 2 Upscale, Mixed Media, AI Motion Designer, Collab, Games, Reference, Extension, ChatGPT Plugin, Academy, Originals, Contests, Plugins, Community.

The most competition-relevant of these, ranked by (recognisability × rebuild feasibility): Genjutsu/Recast (character swap), Angles/Shots/Zooms (camera), Effects/Camera Controls (VFX), ClipCut/Urban Cuts (edit), Relight/Outfit Swap/Style Snap (edit), Popcorn (storyboard consistency), Virality Predictor (novel, and a plausible "Most Viral"/"Most Innovative" play).

No official popularity ranking exists. Third-party inference (weak evidence) puts Cinema Studio #1, Popcorn #2, Genjutsu #3. Higgsfield's own homepage hero — their real prioritisation signal — features Playable, Content Factory, Kling 4.0, Genjutsu Lite, Genjutsu, Seedance 2.5, Supercomputer, Cinema Studio 4.0, MCP for Claude and AI Influencer. Two of the three apps Comfy names are in Higgsfield's own hero rotation, which is a good independent signal that Comfy picked the right targets.

Higgsfield's API and open-source posture

• Higgsfield has a real, self-serve developer API — launched 2026-09-16. Docs at docs.higgsfield.ai (with an llms.txt index), base URL https://api.higgsfield.ai, console at console.higgsfield.ai, plus "Open Higgsfield" at open.higgsfield.ai. Auth is an Authorization header; generation submissions accept an optional Idempotency-Key. The lifecycle is async: submit JSON → get a request_id → poll status_url or receive a webhook → download on completed. SDKs: Python higgsfield-ai/higgsfield-client (Apache-2.0) and TypeScript higgsfield-ai/higgsfield-js (MIT).
• The catalogue is 50+ models, mixing Higgsfield's own (Soul 2, Soul Cinema, DoP, Marketing Studio Image, Soul ID, Genjutsu) with third-party (Kling 2.5/2.6/3.0, Seedance 2.0/2.5, Wan 3.0/3.0 Prime, MiniMax H3, LTX 2.5 Pro, PixVerse 6, Happy Horse 1.1, Recraft, Ideogram, Grok). Published starting rates: Soul 2 / Soul Cinema $0.0032/image, Marketing Studio Image $0.0059/image, DoP $0.125/generation, Kling 2.5 $0.042/s, Seedance 2.5 $0.0738/s, MiniMax H3 $0.13/s, LTX 2.5 Pro $0.17/s, Wan 3.0 $0.20/s. Billing is pay-as-you-go USD with no subscription at any level; failed requests auto-refund; balance expires 1 year from deposit. "Unlimited" model access applies only to the web app — API, MCP, CLI, Canvas and Supercomputer always deduct credits.
• Higgsfield did not open-source its platform, and I found no evidence it ever promised to. The competition's premise is unsupported by any source located. What is open source under the verified github.com/higgsfield-ai org (11 public repos) is tooling, not models:

The higgsfield repo is the actual subject of the September 2026 viral "Mashrabov open-sourced his startup" claim — an LLM training framework, now archived, not the studio, models or workflows.

• The "open source Higgsfield studio" repos that dominate search are third-party, not Higgsfield: Open-Generative-AI by anil-matcha (MIT, 31,380 stars, powered by the Muapi.ai gateway — not Higgsfield), Open-Higgsfield-AI by Autom8AI (a fork, rebranded MATAI), open-higgsfield by wide-trace (Next.js, 38 models, 4,607 stars, connects to the paid Higgsfield API), and thehiggsfieldai/OpenHiggsfield (independent BYOK studio, explicitly "Not affiliated with, operated by, sponsored by, or endorsed by Higgsfield.ai").
• Implication for a ComfyUI rebuild: there is no Higgsfield model or inference stack to fork. The only reusable Higgsfield assets are thin API clients (SDKs, CLI, agent skills) and an archived LLM-training framework. A rebuild is a from-scratch reimplementation of the product surface on top of open models — because every model Higgsfield actually uses (Seedance 2.5, Kling, Veo, Wan 3.0) is closed third-party, and every Higgsfield-branded model (Soul, Soul Cinema, DoP, Speak, Genjutsu, MS Image) is closed and proprietary.

ComfyUI building blocks

Cloud vs local vs API

Key credit facts: API jobs draw from the same monthly credit pool as the Cloud UI — there is no separate API pool. Unused top-ups roll over up to 1 year. All available models can be used for commercial purposes (Comfy's own statement on the pricing page).

The ComfyUI API — concise technical reference

Workflow JSON: two formats. The browser save format (.json) carries node titles, positions, colours, groups. The API format (File → Export Workflow (API)) uses numeric node IDs as keys and omits all layout metadata. Only API format can be submitted. Shape:

{
  "3": {
    "inputs": { "seed": 156680208700286, "steps": 20, "cfg": 8,
                "sampler_name": "euler", "scheduler": "normal", "denoise": 1,
                "model": ["4", 0], "positive": ["6", 0], "negative": ["7", 0],
                "latent_image": ["5", 0] },
    "class_type": "KSampler",
    "_meta": { "title": "KSampler" }
  }
}

Inputs are literals or [node_id, output_index] links. _meta.title is cosmetic. Convert by loading a save-format file in the frontend and re-exporting as API.

Server routes (server.py, search @routes):

Submission payload with a partner node: add extra_data.api_key_comfy_org to the POST body:

{ "prompt": { … }, "extra_data": { "api_key_comfy_org": "comfyui-…" } }

WebSocket message types (/ws): status (exec_info.queue_remaining), execution_start (prompt_id), execution_cached (nodes), executing (node, prompt_id; node: null = done), progress (node, value, max), executed (node, output — only when a node returns a ui dict, not on every completion), execution_success (prompt_id, timestamp), execution_error, execution_interrupted. Custom message types can be registered with api.addEventListener("my.custom.message", handler) and sent with PromptServer.instance.send_sync(...).

Recommended pattern: submit to /prompt, follow /ws until execution_success, then GET /history/{prompt_id} and fetch files via /view. (The docs' own basic_api_example.py is submit-and-forget; websockets_api_example.py is the recommended WebSocket + history pattern.)

Comfy API v2 (the managed path): POST /api/v2/jobs with { workflow, metadata }; GET /api/v2/jobs?metadata[key]=value (paginated, limit ≤500, next_cursor). Idempotency is reject-on-duplicate, not record-and-replay: a reused Idempotency-Key fails 422 idempotency_key_reuse; recovery is to look up the job, not resubmit. Keys expire after 24 h. A naive retry loop around submit() is a double-bill. Self-hosted ComfyUI serves v2 via comfy-api-proxy on port 8189.

Node packs and models relevant to video

Native in ComfyUI core (no custom pack needed):

Custom node packs / third-party:

Partner vs open — the decisive distinction. Comfy's built-in node docs list 322 partner nodes. Confirmed partner (API) nodes include MiniMax H3 (MinimaxHailuo03*), LTX 2.5 (LtxApi25*), FLUX 3 (Flux3*), Wan 2.7 / 3.0 (Wan2*Api, Wan3*Api), plus BFL, Bria, Beeble, Anthropic, Google and others. Using any of these in the critical path disqualifies an entry from Open Weights Only — even where an open-weights sibling of the same model exists, because the node is the closed API. MiniMax H3 and LTX 2.5 are the two cases where you can choose the open-weights route instead. Partner nodes can be disabled entirely at launch with --disable-partner-nodes (or --offline) — a clean way to prove an entry is open-weights-only.

Note the irony: Higgsfield is itself a ComfyUI partner-node provider. The partner list includes Higgsfield alongside Kling, Luma, Runway, Veo, Vidu, Seedance, HeyGen, Topaz and Magnific. Comfy already resells the company whose apps the competition asks you to rebuild.

The licence trap — "open weights" ≠ "open licence"

This is the most under-appreciated risk in the competition, and it is not mentioned on the event page at all. The Open Weights Only category is defined by node type ("no partner nodes or closed-source APIs"), but several models that pass that test carry licences that are revenue-capped, territory-excluded, or non-commercial. For a US-based commercial entrant, two of the three headline open video models have a licensing problem.

Practical consequence: if the entry is intended to be commercially usable by a US company, the only cleanly licensed open video substrate is Wan 2.1 / 2.2 (Apache-2.0). LTX is acceptable under $10M revenue. MiniMax H3 local use is not licensed for a US entity by default — which means the "Built with MiniMax H3" bonus realistically means the hosted API path, and therefore forfeits Open Weights Only. That is the opposite of what the dual open-weights/partner-node status first suggests.

Is there an official app-building SDK?

Yes — but it is three clonable reference apps plus the SDK, not a single "app framework". This is the most actionable finding in this section.

There is no "workflow-as-app" one-click publisher. The Hub's three apps are hand-built Next.js front-ends. The intended path is: clone a reference app → point it at your workflow or Router model → deploy. comfy-cli's comfy build/comfy deploy is the deployment path; the SDK is the runtime client.

What a winning entry looks like

Evidence-based read, with the reasoning exposed.

Which category is most winnable

1. Open Weights Only (RTX 5090) — the most winnable, and the most defensible. It is the only category with a stated, verifiable, binary test ("every model, node and dependency is open-weights or open-source, with no closed API model in the critical path; rented GPUs still count"). It is also the category most competitors will fail by accident, because the two easiest reference apps (comfy-reshoot, comfy-openjutsu) both ship with a closed model in the critical path — comfy-openjutsu calls GPT Image 2.5 through Router for its anchor frame. Dropping that one call is a one-line change that converts a reference app into an Open-Weights-Only entry. The open stack is genuinely available: SCAIL-2 (character swap), Wan Animate 2, Wan 2.2 VACE/Fun-Control, LTX 2.5 + CrossView-Warp (re-shoot), MiniMax H3 local int8, FLUX 3 Image Klein, SAM 3.1, MoGe-2.

2. VFX Workflow Built with LTX ($2,000 cash) — the best risk-adjusted cash prize. $2,000 cash, no hardware logistics, and the requirement is narrow and checkable: LTX must be the core video model. LTX 2.5 is open weights, so this bonus stacks with Open Weights Only. The re-shoot path (CrossView-Warp is an LTX 2.3 IC-LoRA) is already LTX-centric. A VFX workflow built on LTX + CrossView-Warp can plausibly contend for Best Overall, Open Weights Only and the LTX bonus simultaneously.

3. Best Overall ($10,000) — highest value, lowest predictability. No criterion beyond "best overall build", judged by the Comfy team, with no rubric. Every entry is auto-considered. Winning this is a taste call. The realistic route is to win a category and be the most complete, most polished, most obviously-working build.

4. Most Innovative (RTX 5090) — the only category with an explicit argumentative burden. "An app that's never been done, novel custom node packs or uses of ComfyUI. Demo video must explain why the submission is innovative." The obvious rebuilds are taken (Freefield has ~95 tools; comfy-openjutsu has 53 stars). Innovation here means either a novel custom node pack (e.g. a free-orbit camera node that goes beyond WanCameraEmbedding's nine presets, or a keyframe-interpolating camera trajectory node) or an unexpected use of ComfyUI. This is the highest-variance category.

5. Most Viral (RTX 5090) — a distribution contest, not an engineering one. "Delightful to use and made to be shared." The social post is already mandatory, so the distribution mechanic is built in. This rewards a genuinely one-click, no-setup, instantly-gratifying app — which is in tension with the 96 GB-GPU deployments the reference video apps need.

6. Built with FLUX / Built with MiniMax H3 — neither is a clean Open-Weights-Only play. Both bonuses pay API credits, which is itself a signal that they point at the hosted path. FLUX 3 video is Cloud/API-only today, so the FLUX bonus cannot coexist with Open Weights Only. MiniMax H3 looks compatible because it ships open weights — but its community licence excludes the United States as a territory and requires a paid licence (resold only by Comfy, from $5,000/month) for local commercial use. For a US entrant, the H3 bonus realistically means the hosted API, which forfeits Open Weights Only. The only bonus that stacks cleanly with Open Weights Only is LTX ($2,000 cash) — because LTX ships open weights, the prize is cash, and the LTX licence is acceptable under $10M revenue.

What the entry must demonstrate

1. A working app, not a workflow export. The brief explicitly distinguishes "a working creative app with a usable interface and example outputs" from a workflow export. The reference apps are Next.js front-ends with a real UI.
2. A public repo with an OSI licence, the workflow files, and local setup instructions that actually work. The judges may try to run it. comfy-reshoot's README is the standard to match: exact git clone / .env.local / npm ci / npm run dev, plus the comfy build/comfy deploy sequence.
3. A ≤2-minute demo video showing the app running and explaining how it was built. Not a montage. The FAQ is explicit that the work is judged, not the editing — but the video is the only artefact guaranteed to be watched.
4. A live link, if you can afford it. Judges will run it 5–10 times. Budget for that or omit it. A dead link is worse than no link.
5. The mandatory social post tagging #OpenSourceHiggsfield.
6. For Open Weights Only: a stack-openness story you can defend line by line. Every model, node and dependency. Rented GPUs are fine; closed APIs are not. The form asks you to declare this, so an overclaim is a disqualification risk.
7. For Most Innovative: an explicit innovation argument in the video. Name the novel node pack or the novel use, and show why it is new.

The strategic read

The highest-expected-value play is a single entry that is simultaneously (a) Open-Weights-Only clean, (b) LTX-core, and (c) a genuinely usable app — because those three are mutually compatible, the criteria are checkable, and the prize stack is $10,000 + RTX 5090 + $2,000. The natural target is camera re-shoot built on LTX 2.3/2.5 + CrossView-Warp, or character swap built on SCAIL-2 / Wan-Animate-2 with the closed anchor removed. Both are rebuilds of apps the event page itself names, both have open engines, and both have a reference app to clone.

But the licence axis cuts across that plan, and it is the thing almost nobody will check. If the deliverable is meant to be commercially usable by a US company, the substrate choice is constrained:

• Wan 2.1 / 2.2 (Apache-2.0) is the only fully unencumbered open video family — no revenue cap, no territory exclusion, no attribution clause. Wan-Animate-2 (Apache-2.0, native, no pose extraction, text-driven viewpoint control) is the safest character-swap engine, and Wan Fun-Camera / CameraReference3D the safest camera path.
• LTX 2.3 / 2.5 is free for commercial use only under $10M revenue, and 2.5's weights are gated. Fine for this entry; a constraint on the product.
• MiniMax H3 is the trap: open weights, but the US is an excluded territory and local commercial use needs a $5,000/month licence. Do not build the flagship entry on local H3.
• FLUX 3 Image Klein is image-only and revenue-capped; InfiniteYou is non-commercial outright.

So the two candidate architectures are: (1) LTX-core — best for the LTX cash bonus and the CrossView-Warp re-shoot path, accepting a revenue-capped licence; or (2) Wan-core — best for licence cleanliness and the Open Weights Only claim, using Wan-Animate-2 for swap and Fun-Camera/CameraReference3D for camera. The Wan path is the more defensible entry if the goal is a genuinely reusable open asset rather than a prize-maximising one.

The differentiation cannot be "we rebuilt Genjutsu" — Freefield and comfy-openjutsu already did. It has to be one of: a better interface (the reference apps' orbit/keyframe UX is the hardest part and the most visible), a missing mode (Comfy's Openjutsu does person-swap only; Higgsfield's Genjutsu also does Object Swap), a novel node (free-orbit camera beyond nine presets), or a category the crowd ignored (Effects/Transitions, Relight, Outfit Swap, Virality Predictor).

Time is the binding constraint. ~16 hours remain. A clonable MIT reference app plus a working deployment is the only realistic path to a complete, judged entry. Anything built from scratch will not have a demo video, a repo, and a live link by 09:00 PT.

Sources

Competition — primary

Competition — secondary

Higgsfield

ComfyUI technical

Reference apps and open engines

Unknowns and risks

Unknowns (state these as unknown, do not guess)

1. How many free credits the first 200 signups receive. Not published anywhere I could reach. The only credit figures Comfy states are the $250 runner-up prize and the partner bonuses.
2. Whether the signup credits are restricted to Comfy Cloud GPU time, partner nodes, or both. The signup form says only "where the credits land" (the Comfy account email).
3. Whether the submission form will still accept a response at 08:59 PT on 2026-10-30. It is live now; Typeform can be closed server-side at any moment.
4. Who the judges are, and whether there is a rubric. Only "the Comfy team" is inferable; no names, no weights, no tie-break rule.
5. Whether one entry can win multiple main categories (e.g. Most Innovative and Open Weights Only). Only main+bonus stacking is stated.
6. Team entries. Not addressed. "One submission per person" implies a team could file several, but this is an inference.
7. Eligibility restrictions (age, residency, sanctions, employment). None stated.
8. The "special in-person awards" mentioned on both Luma pages. Undefined, unlisted, and absent from the official rules.
9. The "Creator Feature Terms" referenced by a mandatory agreement checkbox. The link is not a link and comfy.org/creator-feature-terms/ is 404. You are being asked to agree to terms you cannot read.
10. Which model actually renders the final pixels in comfy-reshoot — the app is tagged MiniMax H3 but credits CrossView-Warp (an LTX 2.3 LoRA). Most likely H3 generates and CrossView provides the guide; unconfirmed.
11. Whether FLUX 3 Image Klein satisfies "FLUX 3 must be the core model" for the FLUX bonus. It is a FLUX 3 variant and open weights; the full FLUX 3 video model is not. Judgement call.
12. Higgsfield's underlying models for Genjutsu, Soul, Angles, Recast, Effects, Popcorn, Speak 2.0 and Draw-to-Video. All undisclosed by Higgsfield. The one exception, now confirmed: Cinema Studio 4.0 runs on Seedance 2.5 (ByteDance), per Higgsfield's own help centre.
13. What Higgsfield Speak and Popcorn actually do. RESOLVED. Speak = the lip-sync engine inside Lipsync Studio (Higgsfield Speak 2.0, ≈14 credits/5 s @720p, Plus tier required). Popcorn = a storyboard generator producing up to 8 coherent frames from up to 4 references at 1.5 credits/image. Sourced from Higgsfield's Creator Hub help centre.
14. Whether Higgsfield has a public developer API and docs. RESOLVED. Yes — https://api.higgsfield.ai, docs at docs.higgsfield.ai (with an llms.txt index), console at console.higgsfield.ai, launched 2026-09-16, 50+ models, published per-second/per-image rates, Python and TS SDKs. Still unknown: the live higgsfield.ai/pricing and open.higgsfield.ai/pricing pages are JS-only and unreadable, so all plan/credit figures come from Higgsfield's blog or third-party teardowns.
15. The exact status and authenticity of the cease-and-desist. Reported by multiple secondary sources; the primary Reddit disclosure is JS-blocked to this tooling.
16. Whether the competition will proceed as advertised given (15).
17. Whether Higgsfield ever actually promised to open source its platform. No source located supports the claim. The competition's premise appears to be a misreading of an archived Apache-2.0 LLM-training framework.

Risks, ranked

1. Legal / event-cancellation risk (highest). If the reported C&D is real and escalates, the event could be modified, paused, or the prize terms changed after submission. Comfy has reportedly refused to comply, and the LA Luma page carries the "not affiliated with or endorsed by Higgsfield" disclaimer Comfy cites. Mitigation: submit early, keep the repo and demo video self-contained and independently valuable, and avoid using Higgsfield trademarks or the CEO's likeness in Magieva's own entry or marketing. Note that the event's own premise ("Higgsfield promised to open source") is factually wrong, which weakens the framing but not the build.
2. Time risk (binding). ~16 hours remain. A from-scratch app cannot ship a repo, a ≤2-minute demo, and a live link in time. Mitigation: clone a Comfy-Org reference app and adapt it.
3. Compute risk. Both video reference apps were demonstrated on 96 GB GPUs; comfy-openjutsu explicitly says smaller configurations are untested. SCAIL-2 fp16 is ~32.8 GB and runs on a 24 GB 4090 only with offloading (slow). Mitigation: use fp8/mxfp8/GGUF quantisations, or rent a 96 GB GPU (Comfy API RTX PRO 6000 at $3.49/hr), or choose the LTX/CrossView path which is lighter than H3.
4. Open-Weights-Only disqualification risk. The easiest reference apps contain closed API calls (comfy-openjutsu → GPT Image 2.5 via Router). Declaring "fully open" while shipping a closed call is a stated eligibility failure. Mitigation: audit every node and dependency; remove the anchor step; prefer local Wan 2.2 / LTX 2.5 / SCAIL-2 / Wan-Animate-2; verify with --disable-partner-nodes.
5. Licence risk — the invisible one. "Open weights" is not "open licence". MiniMax H3's community licence excludes the United States as a territory and requires a paid licence (resold only by Comfy, from $5,000/month) for local commercial use; LTX is revenue-capped at $10M; FLUX 3 Image Klein is revenue-capped and image-only; InfiniteYou's weights are CC-BY-NC (non-commercial) despite Apache-2.0 code. None of this is mentioned on the event page, and none of it affects eligibility — but it determines whether the deliverable is actually usable. Mitigation: default to Wan 2.1/2.2 (Apache-2.0) for anything intended to be commercially reusable; read the LICENSE file for SCAIL-2 (MIT vs Apache-2.0 conflict), ReCamMaster and ReActor (both unverified).
6. Bonus incompatibility. FLUX 3 video is Cloud/API-only, so a FLUX bonus entry forfeits Open Weights Only. The H3 bonus pays API credits and local H3 commercial use is unlicensed in the US, so it points at the hosted path too. Only the LTX bonus stacks cleanly with Open Weights Only. Mitigation: decide the category first, then pick the model — not the reverse.
7. Crowded field. Freefield (~95 tools, ~117–120 workflows, MIT, published 10/27) and comfy-openjutsu (53 stars) already occupy the obvious ground. Mitigation: differentiate on interface, a missing mode (Object Swap), a novel node pack, or an ignored category.
8. Live-link risk. Judges will run it 5–10 times. A metered deployment that runs out of credits, or a cold-start that takes minutes, reads as broken. Mitigation: cap spend, keep a warm worker if affordable, or omit the link and rely on the video.
9. Unreadable terms. The mandatory "Creator Feature Terms" agreement is a dead link. Mitigation: assume the broadest reading of the feature-rights grant (Comfy, Runpod, LTX, BFL, MiniMax may feature the work with credit) and decide whether that is acceptable before submitting.
10. Form-friction risk. The submission form requires a social post URL and a demo video URL and a repo URL, all as separate required fields, plus a required "List all generative models used". A missing social post is a hard eligibility failure. Mitigation: post the demo and capture the URL before starting the form.
11. Documentation drift. Several Comfy docs pages are AI-generated (the built-in node pages say so explicitly), and the event page's static HTML omits the entire ruleset. Mitigation: treat the i18n bundle and the Typeform field definitions as the authoritative rules; re-verify the deadline and form status immediately before submitting.