FrameMeld Fast V0.1.4-fast.2

FrameMeld Fast is a drop-in runtime variant with the same CLI/API as the standard build. This release keeps the existing Fast source-relative frame policy and adds optional final-output sharpening.

Final sharpening

• --final-sharpen 0.15 applies the tested light profile.
• --final-sharpen 0 disables sharpening.
• Values from 0 through 1.5 are accepted as host-controlled overrides.
• The filter is a 3x3 luma-only FFmpeg unsharp pass applied during the final encode, after interpolation and motion-blur blending.
• The standalone FrameMeld CLI default remains 0 for backward compatibility. Hosts such as Insight Agent can explicitly opt in to 0.15.

Fast automatic policy

• 60 FPS -> 240 FPS
• 90 FPS -> 270 FPS
• 120 FPS -> 240 FPS
• 144 FPS -> 288 FPS
• 180 FPS -> 360 FPS
• 240 FPS and above stay native
• Other sources from 56 FPS to below 240 FPS target exactly 240 FPS
• Sources below 56 FPS retain the smallest integer multiplier that reaches at least 200 FPS

The 270->60 and 288->60 paths continue to use phase-aware sampling. Fractional source rates retain their exact rational timeline.

Installation

Replace the entire framemeld-runtime directory. Do not replace only ffmpeg.exe, because the frame engine and policy are stored under tools/ and lib/.

Verification

• 75 unit tests
• Full RIFE, SVP, deduplication, and motion-blur engine matrix
• Runtime dry-run validation for sharpening disabled, 0.15, and invalid values
• Windows runtime smoke test

SHA-256: 1A24814AEE08ECD7B3CD823E7CABF7F597D15F21CCFB61BF545BD91769377D21

FrameMeld Fast V0.1.4-fast.1

FrameMeld Fast is a drop-in runtime variant with the same CLI/API as the standard build. Insight Agent does not require any code changes.

Fast automatic policy

• 60 FPS -> 240 FPS
• 90 FPS -> 270 FPS
• 120 FPS -> 240 FPS
• 144 FPS -> 288 FPS
• 180 FPS -> 360 FPS
• 240 FPS and above stay native
• Other sources from 56 FPS to below 240 FPS target exactly 240 FPS
• Sources below 56 FPS retain the smallest integer multiplier that reaches at least 200 FPS

The 270->60 and 288->60 paths use phase-aware sampling. Fractional source rates retain their exact rational timeline.

Installation

Replace the entire framemeld-runtime directory. Do not replace only ffmpeg.exe, because the policy and frame engine are stored under tools/ and lib/.

The capability response should report:

• build_flavor: fast
• policy_id: source-relative-fast-v1

Verification

• 70 unit tests
• Full RIFE, SVP, deduplication, and motion-blur engine matrix
• Real 90->270->60 and 144->288->60 phase-aware pipeline tests
• HEVC NVENC smoke test

For the tested 1920x1080 60 FPS sample, the original 300 FPS / 5-tap export took 72.766 seconds. The Fast 240 FPS / 5-tap export took 51.297 seconds: approximately 1.42x faster and 29.5% less elapsed time. Results vary by source and hardware.

SHA-256: BA8551677ABF55632102DE1A66C3F75446765A6C9BDE9FD6620E4CE31A833514

FrameMeld V0.1.3

FrameMeld V0.1.3 adds host-controlled, runtime-verified RIFE GPU selection for hybrid-GPU systems while preserving ordinary FFmpeg pass-through behavior.

Highlights

• Added device-inventory-json-v1, allowing hosts to query a bounded Vulkan inventory without starting an export.
• Added independent --gpu INDEX selection for RIFE/ncnn; this remains separate from NVENC's -gpu encoder selector.
• Added --host-rife-adapter-json so the planned RIFE adapter is not conflated with the planned AMF/QSV/NVENC encoder adapter.
• Added rife_binding structured events based on the device banner actually emitted by ncnn at runtime.
• Explicitly keeps FFmpeg Vulkan and RIFE/ncnn Vulkan as different index spaces. Candidate mappings are not reported as exact until runtime evidence is available.
• Enables host integrations to persist only successful, runtime-verified device mappings and safely re-probe after runtime, driver, or GPU-topology changes.

Compatibility

• Commands without -framemeld continue to pass through to the bundled FFmpeg 9 core.
• Older host integrations remain compatible and continue to use FrameMeld's default RIFE device path.
• AMF and QSV encoder binding remains honestly reported as system-default unless a future runtime can verify the actual adapter.

Download

FrameMeld_ffmpeg-V0.1.3-windows-x64.zip

SHA-256: 9D9AA21565C4084C7A3F5E2E85AE30BCA8248279F6D2EB31EE60EAB07D047105

Extract the archive and run the top-level framemeld-runtime\ffmpeg.exe. This is a Windows x64 prebuilt runtime.

Verification

• 68/68 unit tests passed with the bundled Python/VapourSynth runtime.
• 720p60 RIFE motion-blur to HEVC smoke test passed.
• Nine real headless engine paths passed: RIFE interpolation, RIFE motion blur, balanced/adaptive policy, SVP interpolation, old/SVP/RIFE deduplication, and blur/timescale/color processing.
• --device-inventory-json and the new capability flags were verified on NVIDIA RTX 5070 Ti hardware.

FrameMeld is licensed under GPL-3.0-only. The archive includes the full license, third-party notices, and pinned runtime manifest.

FrameMeld V0.1.2

FrameMeld V0.1.2 adds precise device and export diagnostics for host integrations while preserving ordinary FFmpeg pass-through and standalone encoder fallback behavior.

Highlights

• Added opt-in structured-status-json-v1 lifecycle events covering progress, first frame, first encoded packet, performance summary, child return codes, and failure domains.
• Added device-diagnostics-json-v1 with FFmpeg Vulkan inventory and conservative DXGI-to-RIFE mapping evidence. Mappings without stable UUID/LUID identity are reported as candidate, ambiguous, unmatched, or unknown rather than exact.
• AMD AMF and Intel QSV are reported as separate system-default binding branches; NVIDIA NVENC retains explicit -gpu binding support.
• Host-planned adapter metadata is kept separate from runtime-observed device binding, avoiding false claims that a requested GPU was actually selected.
• Ordinary commands without -framemeld continue to pass through to the bundled FFmpeg 9 core.

Download

FrameMeld_ffmpeg-V0.1.2-windows-x64.zip

SHA-256: 3BECBD156A9FBABCD0EE42C081B5FD6122A279DB2FA08D033E793393382CCDB3

Extract the archive and run the top-level framemeld-runtime\ffmpeg.exe. This is a Windows x64 prebuilt runtime.

Verification

• 65/65 unit tests passed with the bundled Python/VapourSynth runtime.
• 720p60 RIFE motion-blur to HEVC smoke test passed.
• Nine real headless engine paths passed: RIFE interpolation, RIFE motion blur, balanced/adaptive policy, SVP interpolation, old/SVP/RIFE deduplication, and blur/timescale/color processing.
• NVIDIA RTX 5070 Ti was physically tested. AMD AMF and Intel QSV logic is covered by tests but still requires vendor-hardware field validation.

FrameMeld is licensed under GPL-3.0-only. The archive includes the full license, third-party notices, and pinned runtime manifest.

FrameMeld V0.1.1

FrameMeld V0.1.1 upgrades the bundled runtime to FFmpeg 9.0 and completes the host-managed export fallback integration.

Highlights

• Bundled FFmpeg upgraded to 9.0 full build.
• H.264 NVENC, AMF, QSV, and libx264 outputs use High Profile with 8-bit YUV 4:2:0.
• RIFE model loading now works when the FrameMeld runtime is installed under a Unicode path.
• Hosts can opt into host-managed-encoder-fallback so the first hardware encoder failure is preserved instead of being silently converted to a CPU export inside FrameMeld.
• Standalone FrameMeld usage retains the existing automatic same-codec software fallback.

Download

FrameMeld_ffmpeg-V0.1.1-windows-x64.zip

SHA-256: 19424A45814EDD47ECD95D422F3D4ECB40322A7E14A8D28FF3DA7925EDDB1787

Extract the archive and run the top-level ffmpeg.exe. This is a Windows x64 prebuilt runtime.

Verification

• 55 unit tests passed.
• Unicode-path RIFE interpolation passed.
• 1080p60 NVENC export passed and the output metadata reports h264_nvenc.
• Host-managed hardware failure to libx264 fallback passed with a real media sample.

FrameMeld is licensed under GPL-3.0-only. The archive includes the full license and third-party notices.

FrameMeld V0.0.1

FrameMeld 的首个公开预览版本。

主要内容

• 普通 FFmpeg/FFprobe 命令保持原样透传。
• 使用 ffmpeg.exe -framemeld 启动 RIFE 补帧、重复帧修复与运动模糊管线。
• 提供 org.framemeld.cli API 1 能力探测协议。
• 包含已完成主观确认的自动帧率与连续 5→7 抽头模糊策略。
• 支持 NVIDIA、AMD、Intel 硬件编码探测及同编码族软件回退。

下载

FrameMeld_ffmpeg-V0.0.1-windows-x64.zip

SHA-256：8693446ab945dcc516e0041d243bd9b2850ee84c8343c54139edef79b4a74237

解压后使用顶层 ffmpeg.exe。本版本为 Windows x64 预构建运行时。

验证

• 49 个单元测试通过
• HEVC 冒烟导出通过
• RIFE、SVP、三种去重路径、运动模糊、变速和颜色处理验证通过

FrameMeld 使用 GPL-3.0-only；包内包含完整许可证和第三方声明。