luoji12103/2610-superresolution-opt014

VOSR OPT-014 — private course artifact

本仓库保存课程合作使用的 OPT-014 TensorRT 引擎。它沿用既有 VOSR2 模型权重，配合优化后的分块、缩放和颜色处理代码；没有重新训练模型。对应代码在私有 课程仓库，固定源码版本 22edbcaf06fd13c404bffbee3131873ed5a175a4。

Files and runtime contract

• engine.plan: 3,316,011,468 bytes; retained TensorRT engine, with fixed noise42 embedded.
• runtime.json: input/output, precision, environment and source references.
• LICENSE-VOSR / NOTICE.md: upstream license and component attribution.
• Runtime source and environment instructions: OPT-014 package.

Item	Retained setting
GPU/build target	RTX4090, SM89; TensorRT10.13.3.9
Native input/output	RGB8 1920×1080 → RGB8 3840×2160
Engine binding shapes	FP16 NCHW (1,3,512,512) input/output
Tiling	512 pixels, overlap64, two concurrent CUDA graph slots
Model precision/noise	FP16; fixed42 embedded, not a runtime seed option
Processing	GPU bicubic, Gaussian tile blend, retained GPU color kernels
Environment	Python3.11, CUDA12.8, Torch2.9.1, torchvision0.24.1, Triton3.5.1, NumPy2.4.6, Pillow12.0.0

TensorRT plans depend on their GPU/runtime build. Other GPU models, TensorRT versions and fresh installations are not presumed compatible. The runtime does not silently substitute a model or rebuild an engine.

Obtain and use

Download engine.plan from this private repository after authenticating with an account that has access. The course asset index records the repository ID and pinned HF revision after upload. Place the file outside Git, for example local-data/opt014/engine.plan.

After cloning the private course repository, check out source commit 22edbcaf06fd13c404bffbee3131873ed5a175a4 and follow its dedicated environment instructions. From the course repository root:

export VOSR_OPT_ENGINE="$PWD/local-data/opt014/engine.plan"
cd research/vosr-opt014
CUDA_VISIBLE_DEVICES=0 python -m vosr_opt014 \
  --engine "$VOSR_OPT_ENGINE" --input /path/to/frame.png --output /path/to/new-frame-4k.png

The output directory must exist. Inputs must already decode to RGB 1920×1080; no implicit conversion or resize is performed by the CLI. The course exploration adapter separately preserves its registered ×4/other-size adaptation; that is distinct from the native standalone contract.

Evidence and limits

The retained native OPT-014 implementation has historical accepted output/timing evidence. This distribution packages the existing numerical operations with explicit model paths and package-relative imports. Syntax, static dependencies and the retained numerical definitions were checked; this packaged initialization/import path has not had a fresh GPU/model execution. No performance or cross-hardware validation is claimed for a new collaborator environment.

OPT-015 is a microbenchmark, and OPT-016 is a separately visually accepted rebuild with changed arithmetic outputs; neither is included here. The course's frozen H* results use official FP32 VOSR with BasicVSR++, which remains a separate reproduction path. OPT-014 itself is a frame-wise backend and adds no temporal model.

Sources

Base model: CSWRY/VOSR. The retained helper source is based on cswry/VOSR. VOSR project materials use Apache-2.0; component-specific terms remain applicable as described in NOTICE.md. No training data, camera footage or evaluation images are distributed.