PARTNERS WANTED

What do better models change?

I’m H.G.O.D. (H.G.O. Dillenberg) and I build playful, AI-native projects. Everything you see in FrontKit Crew is generated live by open-weight models on home hardware. Now I want to find out what happens when the same projects run on the best models out there. I’m looking for partners who can share API keys, token credits or access to state-of-the-art diffusion, video and music models.

Contact me / drop a key My profile: dillenberg.net See the experiment

Running today: 100 % open models, 0 € cloud

A real clip from the game: the wrestler takes down the giant zombie. Keyframe, motion and sound generated locally, 864×480, 4 s.

  1. Director – Qwen 3.5 (Ollama) writes a fresh variant of each scene: camera, gag, three beats, a shout, the chat.
  2. Keyframe – Qwen-Image-Edit 2.1 Turbo composes shop, enemy style and hero from up to three reference images.
  3. Quality gate – a Qwen 3.5 vision model scores every keyframe and rejects static or broken ones.
  4. Video – a distilled image-to-video model (FastH3, 8 steps) animates 4 seconds with sound.
  5. Live reporter – Qwen3 14B writes the on-air commentary and viewer chat for every situation.
  6. Edit – after each game, HyperFrames cuts the night into a shareable 16:9 news report.
99clips generated
~3:45minutes per clip
7.5/10avg. keyframe score
566reporter lines
190scenes, generated on demand
0 €cloud cost so far

The experiment: jump – or no jump?

Same scenes, same prompts, same references, same seeds where possible. One side runs on my local open-weight stack, the other on your model. Everything goes public, side by side: the clips, the timings, the cost per clip and an honest verdict. If a class-A model makes a huge difference, everyone sees it. If it doesn’t, that’s a result too.

Today · open & local

  • Qwen-Image-Edit 2.1 Turbo
  • FastH3 image-to-video, 864×480
  • Qwen 3.5 / Qwen3 14B
  • One home GPU + a Mac mini
VS
Your model here

Image · edit · video · music · voice · LLM

What I’d measure on a real workload: character consistency across reference images, cartoon style fidelity, motion and comedy timing in 4-second clips, audio, prompt adherence, latency and cost per finished clip.

What I’m looking for

  • API keys or token credits for image generation and image editing with reference images
  • Access to image-to-video models, ideally with native audio
  • Music and voice models for soundtracks and the live reporter
  • Frontier LLMs for directing and writing
  • Or temporary access to bigger GPUs to run larger open models

What you get

  • A public side-by-side comparison on a real, playful production workload
  • Your name and logo as a partner on this page and in the game credits
  • An article on my blog KI ungefiltert and a thread on X
  • Raw data on request: prompts, timings, rejection rates, cost per clip
  • Direct, unfiltered feedback from a builder who ships

Projects that would use your model

FrontKit CrewLive AI game: pick a crew, defend a shop against a zombie horde, watch it as a generated news broadcast. FrontKitProject demo: a service concept for custom-cut protective boards for shop windows, with planner and QR labels. KI ungefiltertMy blog (German) about building with AI – what works, what doesn’t, no hype.

Let’s run the test

Got credits, an API key, early access or GPU time you’d like to see put to work in public? Write to me directly or drop your key via my contact page. The connection is encrypted (HTTPS) and your message goes straight to me.

Contact me on dillenberg.net ▶ 👤 dillenberg.net @jeuner GitHub

Tip: please share a scoped, rate-limited key you can revoke any time. I only use it for the published comparison and delete it afterwards.

Also: Mail gdg [at] dillenberg [dot] net · Telegram @jeuners

FrontKit Crew is a fun project by H.G.O.D. · dillenberg.net · Impressum · Datenschutz