Lost In Space - Video-style-transfer (Experimental test)

This report provides a concise summary of tests performed on two different AMD-based personal computers using the Amuse 3.0.7 video-to-video locomotion model for style transfer, along with system specifications and benchmark results.

**Computer Systems Under Test:**

1. **Minisforum AMD Ryzen 7 255 with Radeon Graphics:**
* **CPU:** AMD Ryzen 7 255 (8 Cores, 16 Threads). It has an integrated Radeon Graphics solution.
* **RAM:** This system is equipped with a substantial **96GB of fast RAM** and 0.50 TB of fast M.2 storage. During Amuse operations, system memory usage was observed to peak at 39GB for an 11-hour run and 50GB for a 6-minute run, with typical usage around 30GB-31GB during high GPU load.
* **Power & Performance:** It operates with a 120-watt power supply, and the CPU's maximum power draw is 53 watts. GPU power draw during tests ranged from 34-41W, with GPU load consistently at 90-99%. Performance can be boosted through BIOS tweaks as official overclocking tools (AMD Adrenalin overclock setting, AMD Ryzen Master series) are unavailable for this model.
* **Software Environment:** Runs Windows 11 with Developer mode, SUDO active, SDK, AMD Adrenalin, WSL, Python V3.13, and Miniconda. AMD Engine 0 handles prompt handling, and AMD Engine 1 manages pipeline preparation and input frames.

2. **Asus AMD Ryzen 9 with Nvidia RTX 3050 (Vivobook M3500QC):**
* **CPU:** AMD Ryzen 9 5900HX (8 Cores, 16 Threads, 3.30 GHz base frequency, 4438 MHz max frequency).
* **GPU:** Features an **NVIDIA GeForce RTX 3050 Laptop GPU** with 4.00 GB of device memory, although AMD Radeon(TM) Graphics is also listed for Vulkan information.
* **RAM:** It has 15.40 GB of DDR4 SDRAM.
* **Operating System:** Microsoft Windows 11 Home (64-bit).

**AMD Amuse Video-to-Video Style Transfer Tests (on Minisforum PC):**

The tests utilised Amuse 3.0.7 with a "Locomotion" model, applying various other models like "ToonYou," "Samaritan3D," "Realistic Vision," and "CyberRealistic". Key parameters influencing results included:

* **Prompt Usage:** While some tests used no prompt, adding specific prompts like "cartoon" or "Disney animation style" influenced the output characteristics.
* **HQ Setting and Resolution:** Turning the "HQ" (High Quality) setting on and increasing the output resolution (e.g., from 480x480 to 896x896) resulted in a slightly better quality rating (5/10 to 5+/10) for the ToonYou model, without affecting the duration in a specific comparative test.
* **Strength Parameter:** This setting dictates the "mix between AI and original video". For the "Realistic Vision" model, a lower strength of 0.20 consistently yielded **high output quality (9/10) with "very good" background and people generation**, while a higher strength of 0.50 resulted in "hopeless" quality. This suggests a critical role of this parameter in achieving desired results.
* **Inference Steps & Duration:** The number of inference steps significantly impacts processing time. For example, in the "Realistic Vision" model, 154 steps took 441 seconds for a 10-second clip, while 34 steps took 644 seconds, indicating that fewer steps don't always mean faster processing, possibly due to other background processes or pipeline inefficiencies. Long clips with high frame counts and step counts, such as a 6-minute clip with 9000 frames and 9000 steps, dramatically increased duration and RAM usage, peaking at 50GB RAM.
* **Model Performance:**
* **ToonYou:** Achieved high output quality, often 9/10 or 10/10, with "good" to "excellent" background and people quality, especially with prompts like "Disney animation style".
* **Samaritan3D:** Generally yielded a 7/10 quality rating. One instance was rated "unusable" despite the numerical score, indicating potential stylistic issues.
* **Realistic Vision:** Showed varied results depending on parameters, capable of achieving 9/10 quality with optimal settings (low strength), but performing poorly otherwise.
* **CyberRealistic:** Yielded moderate quality (6-6.5/10) with poor people generation in initial tests.
* **Background Multi-tasking:** The sources note that background desktop multi-tasking can impact test results.

**System Benchmarks:**

* **Asus AMD Ryzen 9 5900HX:**
* Geekbench 6.4.0: **Single-Core Score of 1979** and a **Multi-Core Score of 7410**.
* OpenCL Score (NVIDIA GeForce RTX 3050 Laptop GPU): **48626**.
* Vulkan Score (AMD Radeon(TM) Graphics): **16991**.
* Cinebench R23: **Multi-Core Score of 1411 pts** and a Single-Core Score of 1448 pts.

Comparing these systems and their performance in the Amuse tests is like comparing two different artists working on digital art. The **Minisforum AMD Ryzen 7** is an artist with a vast canvas (96GB RAM) and a powerful, versatile integrated drawing tablet (Radeon Graphics) that performs best when fine-tuned. The **Asus AMD Ryzen 9** is an artist with a smaller but still capable canvas (15.40GB RAM) and a specialized, high-performance external graphics card (NVIDIA RTX 3050) for specific rendering tasks. Both can produce impressive artwork, but their strengths and optimal working methods differ, with the Minisforum showing greater capacity for intensive, long-duration tasks and the Asus offering strong dedicated GPU performance as demonstrated by its OpenCL score.