Was ich aus lokaler KI rausgeholt habe

Summary

The Morpheus demonstrates how a personal computer can run language, image-generation, image-training, and text-to-speech models locally, without sending prompts or files to a cloud service. He uses them to build a small browser-based visual novel, Just Cancel Now. The project tests what local AI can do—and where human direction and editing are still needed.

Computer and Local Setup

The test PC, supplied by Alternate, contains a Ryzen 7 9700X, an AMD Radeon AI Pro R9700 with 32 GB of graphics memory, 64 GB of RAM, and a Samsung 9100 Pro SSD. The presenter installed Linux and used AMD’s ROCm software stack. Although the distribution was not on AMD’s official compatibility list, the AI workloads ran on the graphics card without manual driver work. Python package conflicts caused some installation friction.

The models and project files stayed on the PC, and the finished game works offline. The presenter estimates that the project used about €2 worth of electricity.

Code Generation

For coding, the presenter used an open-weight Qwen model with 27 billion parameters, compressed through 4-bit quantization to fit in graphics memory. Quantization substantially reduces model size, with some loss of accuracy; he describes 4-bit as a useful compromise.

The model produced a working browser-game excerpt—seven files and nearly 700 lines of code—in just over five minutes. Its context window was set to about 65,000 tokens, but large, open-ended tasks could fill it quickly. The presenter recommends dividing work into small, explicit tasks and specifying which files to change.

The model sometimes got stuck troubleshooting without making edits, but a clearer follow-up request resolved an error quickly.

Image Generation and Consistent Characters

The presenter used ComfyUI to generate images with the video’s named image model, rendered in the subtitles as “Crea/Krea 2.” Its Turbo version generates images quickly; the RAW version is used for training. The workflow involves loading the model, encoding the prompt, setting image dimensions, sampling from noise, and saving the result. Generated PNGs can retain the ComfyUI workflow for reuse.

Text prompts alone did not keep characters consistent across images, so the presenter trained small LoRA add-ons for the characters. After an initially poor attempt, he refined the captions and training settings: 400 steps, 768-pixel training images, and a higher LoRA rank. Training a character took about 45 minutes and around 18 GB of graphics memory.

The results were more consistent, though details could still be wrong. For example, a ficus character never acquired a face.

Speech Generation

For dialogue, the presenter used Qwen 3 TTS to generate voices locally. It produced 55 lines—about four and a half minutes of audio—in roughly three minutes.

He found the voices usable for a hobby project, but had to handle performance, volume matching, and silence trimming himself.

What AI Did—and Did Not—Do

The presenter says only about a third of the finished game came directly from Qwen. He wrote the puzzle logic and an initial version of the dialogue, then used the model for interface work, revisions, and brainstorming. Of roughly 180 dialogue suggestions, around one in ten made it into the game.

AI-generated reviews were not a reliable substitute for human testing. Qwen praised the game in five reviews, but the presenter found it unfun when he played it himself. In a screenshot-based test, the model also became stuck in a maintenance terminal and pressed “test” 131 times.

His conclusion is that models can produce code, images, voices, and ideas quickly, but people still need to choose, evaluate, correct, and combine the results.

Performance, Trade-Offs, and Conclusions

The 32 GB graphics card could not hold two of the larger workloads at once, so models had to be unloaded between tasks. The fast SSD helped reload them: Qwen took about 3.5 seconds to load, and the image model about 4 seconds.

The presenter says a single-GPU setup is adequate for one project at a time, while systems with Threadripper Pro CPUs, more memory, and multiple GPUs are better suited to parallel work and larger datasets.

He compares the local language model’s capabilities to roughly the level of cloud models from the previous November, while citing Artificial Analysis as placing it somewhat above models released before March of the current year. In his comparison, cloud use of the same model for the project would have cost about €3–€9, versus roughly €2 in local electricity.

Cost alone does not make local hardware worthwhile for everyone. Other benefits include privacy, avoiding quotas and subscriptions, offline access, and unrestricted experimentation.

Reviews, Guides, and Tutorials Covered

• A hands-on guide to running AI models locally on AMD hardware with ROCm.
• A practical explanation of 4-bit quantization and model memory requirements.
• A ComfyUI image-generation workflow, including reusable workflow metadata in PNGs.
• A trial-and-error LoRA character-training walkthrough, with settings and observed limitations.
• A comparison of AI-generated game reviews with human playtesting, highlighting the models’ shortcomings.

Speaker and Sources

The Morpheus is the sole presenter and narrator. He cites Artificial Analysis for model comparisons. The video is a collaboration with Alternate, AMD, and Samsung.