{
  "id": 5063227,
  "title": "I tried to build an AI anime on a 16GB MacBook. Here is exactly where it broke",
  "url": "https://urgent.news/2026/09/02/i-tried-to-build-an-ai-anime-on-a-16gb-macbook-here-is-exactly-where",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-02T10:05:42.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/morinaga/i-tried-to-build-an-ai-anime-on-a-16gb-macbook-here-is-exactly-where-it-broke-18g3"
  },
  "original_language": "en",
  "account": "I aimed to test whether a 16GB MacBook could create an animated short using AI. The machine in question is an M1 Pro with 16GB of unified memory. Over the course of three days, I measured the performance of various components, including music synthesis, voice synthesis, character art, and video generation. Two aspects proved more successful than anticipated, while one area completely failed, highlighting the hardware's limitations.\n\nFor the music, I attempted to create beat synthesis without purchasing a subscription. I wrote a synthesiser for plucked strings using the Karplus-Strong algorithm. Initially, the result was unlistenable, but upon comparing it to a reference track, I discovered that 85.2% of the energy was in the bass range, and only 1.5% was in the melody range. Additionally, the left-right correlation remained at 1.00, indicating lack of width, even though I tried to pan individual notes. The audio synthesis took only 3.5 seconds for 48 seconds of finished audio.\n\nRegarding voice synthesis, I faced challenges with four different models. The issues stemmed from the TTS model misreading certain period vocabulary, resulting in incorrect pronunciations. For instance, the title \"Akechi, Governor of Hyūga\" was misread as \"Akechi Mamoru.\" This incorrect parsing caused significant problems, and no amount of model swapping could rectify the issue. Checking proper nouns before synthesis using the VOICEVOX endpoint proved crucial. For the voices themselves, Style-Bert-VITS2 with an emotional corpus showed promise, but it required Python 3.11 and struggled with CPU inference due to differences in fp16 and fp32 data types.\n\nCharacter art presented its own set of challenges. While character consistency was maintained using the same seed and identical prompt descriptions, hands proved to be the most difficult aspect to generate consistently. Close-ups of hands were deliberately chosen to work around this issue, as faces and eye-lines carried the dramatic beats effectively. However, negative prompts were insufficient, leading to unwanted elements such as roses growing on a Japanese castle and a naginata transforming into a katana.\n\nCreating video footage was an additional challenge. The Wan 2.1 T2V 1.3B model, run through ComfyUI, struggled to complete even a brief two-second clip in 33 frames at 20 steps. After 90 minutes, only 9 out of 20 steps had been completed, with 23.3GB of swap memory in use, and 19% of system memory free. The text encoder alone consumed 6.7GB of memory, and the total memory required exceeded the 16GB of unified memory available on the laptop. Consequently, the process was not only slow but also non-functional.\n\nSeveral costly mistakes were made during this experiment. Downloading the same model twice from different sources led to the duplication of 21GB of data on the SSD, further exacerbating memory issues. Additionally, using ComfyUI-format weights and calling diffusers.from_pretrained() resulted in another 5.2GB being downloaded, leading to unnecessary disk usage. To avoid these pitfalls, it is essential to move the file instead of copying it, delete the cache entry after downloading, and utilize ComfyUI directly with the existing weights if available.\n\nIn summary, while certain aspects of AI-generated audio and visuals yielded positive results, the limitations of the 16GB MacBook became apparent when attempting to create video content. Music synthesis was feasible with a custom solution, voice synthesis required careful attention to data quality, and character art maintained consistency despite challenges. Video generation, however, was rendered impractical due to insufficient memory resources, highlighting the importance of hardware specifications in AI-driven projects.",
  "summary": "I wanted to find out whether a laptop can make an animated short with AI. Not \"can a model generate a picture\" — the whole thing: music, voices, character art that stays the same person across shots, and motion. The machine is an M1 Pro with 16GB of unified memory. Everything below was measured on it over three days. Two parts worked better than I expected. One part did not work at all, and that…",
  "key_points": [
    "16GB MacBook struggled to generate video content",
    "Music synthesis succeeded with custom solution",
    "Voice synthesis required careful data quality"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}