ENTANGLED - A 3-minute sci-fi short using 100% local open-source models. Complete Technical Breakdown [ Character Consistency | Voiceover | Music | No Lora Style Consistency | & Much More! ]
A 3-minute sci-fi film created entirely with open-source AI models, no cloud services or proprietary tools.
An independent filmmaker has released 'Entangled,' a technically ambitious 3-minute sci-fi short created entirely with local, open-source AI models, bypassing cloud services and proprietary tools like Runway or Pika. The entire pipeline ran on a single RTX 4090 using ComfyUI, with the creator resisting the temptation to use cloud-based 'brute force' solutions. The project represents a significant milestone for decentralized AI filmmaking, proving that complex narrative video can be produced without relying on external APIs or expensive subscriptions.
The technical workflow was meticulously documented, revealing several innovative approaches. Script compression from 7 to 3 minutes was handled by the local Jan app running Qwen 3.5 35B. Character consistency—the project's biggest hurdle—was solved not by training LoRAs but by using Flux.2 Dev (FP8) with highly structured JSON prompts, treating image generation like code. For audio, base voices from Qwen Voice Designer were enhanced with VibeVoice 7B for emotion, then fed into LTX generations for lip-sync. A key pro-tip was creating a complete animatic with stills and voiceover before generating any video, ensuring precise pacing and minimizing wasted compute.
- Entire 3-minute film created with 100% local open-source models on RTX 4090, using ComfyUI
- Solved character consistency without LoRAs using Flux.2 Dev and structured JSON 'code-like' prompts
- Workflow included Qwen 3.5 35B for script, VibeVoice 7B for audio, and animatic-first approach for pacing
Why It Matters
Proves professional-grade AI video is achievable locally, reducing costs and increasing creative control for indie creators.