ComfyUI-SaveVideoHQ

A ComfyUI custom node for saving VIDEO outputs at professional quality tiers — ProRes ladder, All-Intra H.264 / H.265, and mathematically lossless FFV1 — with an honest 10-bit pipeline that preserves the diffusion model's float precision instead of throwing it away in an uint8 cast.

Built for video diffusion workflows (LTX-Video, Wan, Hunyuan, etc.) where particle-heavy or gradient-heavy output gets mangled by ComfyUI's default SaveVideo and the tune=film defaults of most encoders.

Why this exists

ComfyUI's built-in SaveVideo has no rate-control / GOP / pix_fmt knobs. VHS_VideoCombine exposes basic controls but doesn't ship sane "I want a ProRes HQ master" or "All-I H.264 4:4:4" presets, and uses a uint8 RGB pipeline that quantizes 10-bit pix_fmts to 8 bits before encoding.

This node:

• Ships 19 named professional presets (ProRes Proxy → 4444 XQ, H.264 / H.265 All-I in 4:2:0 / 4:2:2 / 4:4:4, FFV1 lossless) with consistent naming <codec>_AllI_<subsampling>_<bitdepth>_<quality>
• Uses a 16-bit RGB intermediate (uint16, rgb48le) when the target pix_fmt is 10-bit so float precision actually reaches the encoder
• Tunes H.265 with tune=grain:deblock=-3,-3:no-sao so particles and high-frequency detail aren't smeared by HEVC's aggressive default filters
• Is intra-only for every preset — inter-frame compression mangles particles regardless of bitrate; All-Intra is the only sane choice for diffusion output going into post-production
• Has a Custom preset for full manual control of codec / CRF / bitrate / GOP / preset speed / pix_fmt

Install

Via ComfyUI Manager (recommended once registered)

1. Open ComfyUI Manager
2. Click "Install Custom Nodes"
3. Search "Save Video HQ"
4. Install, restart ComfyUI

Via git clone

cd ComfyUI/custom_nodes
git clone https://github.com/xergon/ComfyUI-SaveVideoHQ
cd ComfyUI-SaveVideoHQ
pip install -r requirements.txt

Restart ComfyUI.

Manual single-file install

Drop save_video_hq.py into ComfyUI/custom_nodes/. ComfyUI bundles av (PyAV) so no extra pip install is needed in most installs.

Usage

Right-click the canvas → Add Node → video → Save Video (High Quality), then connect a VIDEO output (from CreateVideo, LTX-2.3 FLF2V, etc.) to the node's video input.

[your video diffusion graph]
        |
     VIDEO
        |
        v
+------------------------------+
| Save Video (High Quality)   |
|                              |
| preset = ProRes_HQ_4:2:2_... |
| filename_prefix = render     |
+------------------------------+
        |
        v
   ComfyUI/output/render_00001.mov

Presets

All presets are All-Intra (every frame a keyframe). Inter-frame prediction is disabled because it mangles particle / high-frequency content regardless of bitrate.

ProRes (intra-only by design, MOV container)

| Preset | Subsampling | Bit depth | ~Bitrate | Use case | |---|---|---|---|---| | ProRes_Proxy_4:2:2_10bit_(~45Mbps) | 4:2:2 | 10-bit | 45 Mbps | Preview / proxy edit | | ProRes_LT_4:2:2_10bit_(~102Mbps) | 4:2:2 | 10-bit | 102 Mbps | Light edit | | ProRes_422_4:2:2_10bit_(~147Mbps) | 4:2:2 | 10-bit | 147 Mbps | Standard edit | | ProRes_HQ_4:2:2_10bit_(~220Mbps) | 4:2:2 | 10-bit | 220 Mbps | Broadcast HQ master | | ProRes_4444_4:4:4_12bit_(~330Mbps) | 4:4:4 | 12-bit | 330 Mbps | Feature / alpha-channel edit | | ProRes_4444_XQ_4:4:4_12bit_(~500Mbps) | 4:4:4 | 12-bit | 500 Mbps | VFX master |

H.264 All-I, 8-bit (broad GPU decode acceleration)

| Preset | Subsampling | CRF | |---|---|---| | H264_AllI_4:2:0_8bit_CRF14 | 4:2:0 | 14 | | H264_AllI_4:2:2_8bit_CRF12 | 4:2:2 | 12 | | H264_AllI_4:4:4_8bit_CRF10 | 4:4:4 | 10 |

H.264 All-I, 10-bit (some GPUs require software decode)

| Preset | Subsampling | CRF | |---|---|---| | H264_AllI_4:2:0_10bit_CRF14 | 4:2:0 | 14 | | H264_AllI_4:2:2_10bit_CRF12 | 4:2:2 | 12 | | H264_AllI_4:4:4_10bit_CRF10 | 4:4:4 | 10 | | H264_AllI_4:4:4_10bit_lossless | 4:4:4 | qp=0 (lossless) |

H.265 All-I (smaller files than H.264 at equivalent quality, ~30-40% size reduction)

All H.265 presets use particle-friendly flags: tune=grain:deblock=-3,-3:no-sao=1.

| Preset | Subsampling | Bit depth | CRF | |---|---|---|---| | H265_AllI_4:2:0_8bit_CRF14 | 4:2:0 | 8-bit | 14 | | H265_AllI_4:2:0_10bit_CRF14 | 4:2:0 | 10-bit | 14 | | H265_AllI_4:2:2_10bit_CRF12 | 4:2:2 | 10-bit | 12 | | H265_AllI_4:4:4_10bit_CRF10 | 4:4:4 | 10-bit | 10 | | H265_AllI_4:4:4_10bit_lossless | 4:4:4 | 10-bit | lossless=1 |

Lossless

| Preset | Codec | Container | |---|---|---| | FFV1_lossless_4:4:4_10bit | FFV1 | MKV |

Custom

Pick Custom to expose manual controls:

• custom_codec: libx264 / libx265 / prores_ks / ffv1
• custom_container: mp4 / mkv / mov
• custom_rate_control: crf / bitrate / lossless
• custom_crf: 0-51 (lower = better quality)
• custom_bitrate_mbps: 1-2000 (used when rate_control=bitrate)
• custom_preset_speed: medium / slow / slower / veryslow / placebo / fast / faster / veryfast / ultrafast
• custom_all_intra: true (GOP=1) / false
• custom_pix_fmt: yuv420p / yuv422p / yuv444p (8-bit) or yuv420p10le / yuv422p10le / yuv444p10le (10-bit)

Bit depth fidelity

Most ComfyUI video nodes do (frames * 255).astype(uint8) before encoding — meaning a 10-bit pix_fmt receives only 8 bits of source data. This wastes file size and produces no quality improvement over 8-bit.

SaveVideoHQ detects 10-bit / 12-bit pix_fmts and switches to a 16-bit RGB pipeline:

arr = (frames.cpu().numpy() * 65535.0).clip(0, 65535).astype("uint16")
frame = av.VideoFrame.from_ndarray(arr_per_frame, format="rgb48le")

The encoder receives 16 bits per channel (more than the diffusion model's BF16 effective precision) and quantizes down to the target 10-bit / 12-bit pix_fmt.

You see this in real output as: less banding in skies / smoke / gradients, cleaner color blending for translucent particles, fewer posterization steps in dark areas.

Why intra-only for diffusion video

Inter-frame compression (P-frames, B-frames) predicts each frame from neighbours and stores residuals. For high-frequency content (particles, grain, sparkles) prediction fails frame-to-frame, residuals are large, and the codec spends bits poorly or the deblocking / SAO filters smear the failed predictions.

Particle-heavy diffusion output looks "fizzy", smeary, or temporally unstable when saved with default H.264 / H.265 inter-frame settings — even at very high bitrates. All-Intra removes this entire failure mode at the cost of ~3-5x larger files.

H.264 vs H.265 for particles (short version)

• H.265 inter-frame: avoid for particles. Default deblock + SAO + motion compensation combine to smear high-frequency detail regardless of bitrate.
• H.264 inter-frame: better than H.265 for particles thanks to mature psy-rd defaults and gentler deblocking, but still inferior to All-I.
• All-Intra (H.264 or H.265): roughly equivalent quality at same bitrate; H.265 produces ~30-40% smaller files at equivalent visual quality.

Acknowledgements

Created by @xergon for high-quality video diffusion workflows.

License

MIT — see LICENSE.