README.md

metadata

title: Celebrity Deathmatch
emoji: 🥊
colorFrom: red
colorTo: yellow
sdk: gradio
sdk_version: 6.0.0
app_file: app.py
pinned: true
short_description: Two photos in, a claymation celebrity death match out
tags:
  - track:wood
  - sponsor:modal
  - sponsor:openbmb
  - achievement:offbrand
  - thousand-token-wood
  - off-brand
  - best-demo
  - text-to-video
  - image-generation
  - text-to-speech
  - gradio
  - modal
models:
  - openbmb/MiniCPM-V-2_6
  - black-forest-labs/FLUX.1-schnell
  - Lightricks/LTX-Video
  - Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign

🥊 Celebrity Deathmatch

Upload two photos. Our AI ring director books the brawl — a claymation fight script, a rendered keyframe reel, a declared winner — then turns it into one continuous fight video with two ring announcers screaming over the action and the crowd going wild.

It's MTV's Celebrity Deathmatch as an AI-native toy: pure spectacle, zero practical value, maximum fun. (That's the Thousand Token Wood track in one sentence.)

⚠️ Parody. Every visual is an AI-generated claymation caricature of a public figure, for comedic effect. Not real. No real people were harmed.

▶️ See it in action

🎬 Watch the 60-second demo · 🥊 Try it live · 🔗 Launch post

Why it's worth a look

• 🎙️ Two-announcer voiceover, not a silent clip. Every beat is called by Nick (dry, sarcastic) and Johnny (loud, over-excited) — two designed voices from Qwen3-TTS VoiceDesign — mixed over a bell, a crowd murmur bed, and a winner roar. The fight has a soundtrack, like the real show.
• 🎨 Off-brand UI. No default Gradio look: a custom claymation-fight art direction — Anton display type, fire-and-clay palette, tale-of-the-tape stat bars, animated winner banner.
• 🧱 Real caricatures, real stakes. MiniCPM reads both photos, invents a fighting persona, signature move, and stat line for each, then choreographs a 5-beat arc and picks a winner.

How it works

A four-model pipeline — every model under the hackathon's 32B cap:

Stage	Role	Model	Params
1	Fight director — reads both photos → fight card JSON	MiniCPM-V-2_6 (OpenBMB)	8B
2	Claymation keyframe reel	FLUX.1-schnell (BFL)	12B
3	Keyframes → continuous fight video (opt-in)	LTX-Video (Lightricks)	2B
3	Two-announcer voiceover	Qwen3-TTS VoiceDesign (Qwen)	1.7B

The entire fight pipeline runs on small models — all ≤12B, well under the 32B cap. No giant foundation model anywhere: a clever chain of small specialists (read → draw → animate → voice) does the whole show. That's the Build Small ethos.

photo A + photo B
   └─▶ Stage 1  MiniCPM-V-2_6 → fight card (fighters, 5 beats, 2-announcer commentary, winner)
        └─▶ Stage 2  FLUX.1-schnell → 5 claymation keyframes
             └─▶ Stage 3 (Animate)  LTX-Video → chained clips
                  + Qwen3-TTS → Nick/Johnny voiceover  + crowd SFX  + burned-in captions
                  → one MP4

The reel (Stages 1–2) is the fast default; Animate (Stage 3) is opt-in because it's the GPU-heavy step.

Tech

• Frontend: this Gradio 6 Space (custom CSS, no template look).
• Backend: two Modal GPU apps — deathmatch (MiniCPM A10G + FLUX L40S + ASGI gateway) and deathmatch-video (ComfyUI + LTX on H100, Qwen3-TTS on A10G).
• Audio/video post: pure ffmpeg — xfade stitch, fit-to-beat TTS mixing, synthesized crowd SFX, burned-in captions.
• Wired to the backend via DEATHMATCH_API_URL (set automatically on deploy).

Sponsors we built on

• 🤗 OpenBMB — MiniCPM-V-2_6. The whole show starts here: one 8B vision-language model reads both fighter photos in a single call and returns the entire fight card as JSON — names, personas, signature moves, stat lines, a 5-beat script, two-announcer commentary, and a winner. It is the ring director.
• ▲ Modal. Every GPU stage runs on Modal — MiniCPM (A10G), FLUX (L40S), and the ComfyUI + LTX-Video + Qwen3-TTS video app (H100 / A10G) — behind warm-pooled @app.cls containers and an ASGI gateway. The HF Space stays CPU-only and calls Modal over HTTP; one modal deploy ships each app.

Run it locally (no GPU)

DEATHMATCH_MOCK=1 python app.py

Mock mode exercises the entire UX on CPU — canned fight card, placeholder keyframe reel, and a stand-in fight video — so you can see the whole flow without burning a single GPU minute.