VELVET VICE LTX — Auto NVENC Video Output

Every LTX workflow ends the same way: you've got a stack of frames, you want an MP4, and someone has to run FFmpeg. VelvetViceLTXAutoVideoCombine is the Velvet Vice pack's take on that step, and its one clever trick is that it tests whether your GPU can actually encode before choosing the encoder, rather than assuming NVENC works and dumping a cryptic error on you two minutes into the render.

How it works

Under the hood it delegates to VideoHelperSuite's VHS_VideoCombine - so this node is a smarter wrapper, not a reimplementation, and it needs VHS installed. The encoder_mode has three settings:

• AUTO - NVIDIA NVENC / CPU H.264 FALLBACK (default): before encoding, it runs a real probe - a 256×256 test frame piped through FFmpeg's h264_nvenc with a short timeout. Probe succeeds? It uses VHS's NVIDIA H.264 preset (video/nvenc_h264-mp4) at your bitrate. Probe fails, or NVENC isn't registered? It falls back to CPU libx264 at your CRF. And if NVENC dies mid-encode in AUTO mode, it retries once on CPU instead of killing the render.
• NVIDIA NVENC ONLY - no probe, no fallback. If NVENC isn't there, it raises immediately so you find out before spending minutes rendering.
• CPU H.264 ONLY - forces libx264, never touches NVENC.

The bitrate input is nvenc_bitrate_mbps_at_24fps, which gets scaled by your actual frame rate (so 60 Mbps at 24 FPS becomes ~80 at 32 FPS). There's also a fast-path patch for FP16 frames: VHS normally converts FP16→RGB8 on the CPU, which is surprisingly slow for multi-gigapixel clips; the wrapper swaps in a faster FP32-based converter, scoped only to its own encode call.

Inputs and outputs that matter

• images - your frame sequence. audio (optional) - mux audio into the MP4.
• frame_rate - 24 default; LTX's native cadence. filename_prefix - video/LTX23-FINAL by default; supports VHS-style %date:...% tokens and is scrubbed Windows-safe.
• encoder_mode, nvenc_bitrate_mbps_at_24fps, cpu_crf - the encoding knobs above. CRF 15 default is visually clean; raise it if file size matters.
• pingpong (loop A→B→A), save_metadata (embed workflow), trim_to_audio (CPU path only - it's passed only when libx264 runs), save_output.
• Output: Filenames of type VHS_FILENAMES - wire it to a save/preview node or the pack's cleanup node. It's marked as an output node.

Installing it

Install the pack as VELVET VICE - LTX (velvet-vice-ltx) via ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-ltx

Restart and hard-refresh with Ctrl+F5. The pack itself has no pip deps, but this node genuinely requires VideoHelperSuite and FFmpeg (VHS bundles an FFmpeg; otherwise it must be on PATH). No VHS, no video.

Troubleshooting

• "VideoHelperSuite is required" - install VHS via ComfyUI Manager, restart.
• NVENC ONLY mode errors at queue time - that's the probe working as designed; your FFmpeg build lacks h264_nvenc (or your card/driver doesn't support it). Switch to AUTO or CPU.
• AUTO mode prints "NVENC fallback reason" - read that line; it's the actual FFmpeg stderr from the probe, usually driver or unsupported-pixel-format trouble.

It's workflow plumbing, but it's well-built plumbing: one encode, never two, with a decision you can actually see in the console. The pack also ships three version-pinned variants of this node (v0.1.9, v0.1.10, v0.1.12) so older workflow files keep their exact behavior - if you're using one of those, read its page instead.