Biniou on a 4 GB Windows GPU: Two Real Video Tests

What we actually tested

Tested October 15, 2026: the actual biniou interface and its native Modelscope video function, using two English prompts. This is a configured local test of one checkpoint, not a benchmark of every biniou module or its untouched one-click installer. The cover is an AI-generated editorial illustration, not a screenshot or generated test frame.

The host ran Windows 11 build 26200, a Ryzen 5 5600H, 27.86 GiB usable physical RAM and an RTX 3050 Laptop GPU with 4096 MiB dedicated memory, driver 616.92. A real CUDA matrix multiplication passed before generation. Other user applications remained open.

Setup and the changes we made

We used biniou commit 643d0b0, local Python 3.11.15, PyTorch 2.1.0+cu121, Diffusers 0.34.0, Transformers 4.51.3 and Gradio 3.50.2. Dependencies were installed into an isolated test environment after inspecting the Windows installer; system Python, global Git settings and drivers were unchanged.

The selected model was damo-vilab/text-to-video-ms-1.7b, now hosted under ali-vilab, at revision 8227dddca75a8561bf858d604cc5dae52b954d01. We explicitly selected its official FP16 safetensors variant and low-memory loading. The 13 model files totaled 3.42 GiB; all three weight files matched the publisher’s SHA256 checksums.

The UI was restricted to loopback HTTP, with no public share; outbound TLS verification remained enabled. We retained biniou’s sampler, Compel conditioning, sequential CPU offloading, UNet chunking and VAE slicing/tiling. Generation used the completed local cache with Hugging Face offline flags, without paid inference or a cloud generation API.

Two exports, measured separately from downloads

Both runs used 256×256, 8 FPS, DDIM, guidance 7.5, seed 12345, one output and GFPGAN disabled. Native-call time includes pipeline loading and export, but excludes imports and downloads. The benchmark called the actual biniou backend; the browser interface was checked separately.

Both H.264 MP4s fully decoded without errors: one second and two seconds respectively, with all expected frames. Prompts and settings both changed, so this does not isolate the effect of step count. More steps alone cannot be credited with the improvement.

What the pictures showed

The first output was technically valid but visually unusable: bright multicolored patterns, an orange rounded area and small frame changes. The requested boat and water scene were absent. Counting unique frames would have hidden that failure.

The second output showed a natural-colored rocky shoreline with changing waves and sea foam. However, the small red subject did not form a recognizable lighthouse. Foreground rocks and framing shifted despite “static camera.” It demonstrated local video generation, but did not meet the complete brief or establish production quality.

We inspected every decoded frame in contact sheets. Subject fidelity and spatial continuity were assessed separately from file validity. No interpolation, replacement footage or cleanup was used. The cause of the first failure was not established by these two runs.

Memory, storage and waiting time

Three-second samples peaked at 5035.6 / 5166.3 MiB process resident RAM and 3095 / 3261 MiB whole-GPU usage for the two trials. Whole-device measurements can include other applications and miss peaks between samples. Private committed memory approached 9.5 GiB; this is distinct from resident RAM. PyTorch allocation counters exceeded dedicated VRAM under Windows WDDM, so we do not report them as physical video-memory use.

Environment files totaled 6.56 GiB, Python 0.14 GiB, package caches 3.41 GiB and source/model files 4.07 GiB, including an unused partial download. These are logical file lengths, not allocated disk blocks. Free disk after testing was 67.48 GiB. The 2.5 GB PyTorch wheel download took about 26½ minutes; the FP16 model download took about 32 minutes while sharing bandwidth. Neither is inference time.

Limits and a useful next step

Optional Triton and deprecation warnings did not stop execution. Zeroscope, AnimateDiff, image-to-video, audio, 3D, higher resolutions, long clips and browser cancellation remain untested. Read the biniou source and installation notes, the exact model card and versioned pipeline documentation. The model card has noncommercial/research terms and differing NC versus NC-ND labels; no unrestricted commercial-use claim follows from this test.

For this configuration, local experimentation worked; fast, reliable promotional production was not established. Start with one short measurable brief, inspect the exported frames and change one factor per follow-up run. Use the Video Prompt Composer to define the shot and the quality-review workflow to judge the result.

Frequently asked questions

Continue with a practical next step

Created by AI Craft Pad. Contact or report an issue. Check the current generator controls and review the output before use.