Official Topaz · 视频插帧 (T8 EXP)

What it does

MiniMaxH3TopazFrameInterpolationEXPT8 runs Topaz's own tvai_fi frame interpolation on a finished clip: 2x or 4x the frame rate, same duration, same resolution, same audio. 24 fps becomes 48, 30 becomes 60. It is not slow motion and it is not an upscale - the pack keeps those as separate nodes, and you chain them serially rather than hoping one does both.

Why bother? H3 outputs 24 fps. Plenty of delivery targets want 60, and a 24→60 resample looks like stutter. Frame interpolation is the companion to a restorer, never a substitute for one - the two are different jobs and comparing them is the standard way these conversations go wrong (upscaling essay).

What you need before it will run

A licensed, installed Topaz Video. The node doesn't ship, download or unlock anything - it calls ffmpeg, tvai_up and tvai_fi out of your real install, and it needs you to have logged in and downloaded the interpolation model inside the Topaz app first. The app itself does not need to be open when you queue.

Because of that, topaz_runtime is required and can't be typed by hand: it's the output of the pack's Official Topaz · 环境检查 node, and wiring anything else raises 请连接 Official Topaz 环境检查节点. That environment node takes three paths - the program directory, the model JSON definitions directory and the model data directory - and they are often in different places on Windows. Read its report_json before you run anything; it lists which model IDs and multipliers actually have candidate weights on your machine. A file being present has never meant a model runs.

The inputs you'll touch

source_video must be an uncropped local file VIDEO. A frame batch gets rejected with a message telling you to save it as a video first - so edits, trims and crops have to be baked into a file before this node sees them.

model_id defaults to apo-8 (Apollo), a solid general-purpose pick that handles non-linear motion well. The list also has apf-2 Apollo Fast, chr-2 Chronos and chf-3 Chronos Fast for general variable-rate work, and aion-1 Aion as the newer general option. If a model isn't in the dropdown, custom_model_id overrides the selection.

multiplier is 2x or 4x. duplicate_threshold defaults to 0.01: 0 or a negative value turns duplicate-frame detection off, and nudging it too high starts deleting genuinely still frames. Leave it where it is unless you can see artifacts.

output_profile decides the container and depth, and it's stricter than you'd expect. The default delivery_h264 accepts 8-bit SDR only; 10-bit SDR needs delivery_hevc_main10, which encodes HEVC Main10. HDR is not supported, and the Topaz path rejects HDR, variable-frame-rate, interlaced and rotation metadata outright rather than quietly converting them. Normalise odd sources before you get here.

The rest is plumbing: output_directory (blank writes to ComfyUI/output/MiniMaxH3-Topaz), vram_fraction (0.8), engine_instances (0 is the stable choice; more instances means more VRAM), gpu_device_index (0 on a single card).

Outputs

interpolated_video is the new clip. source_video passes your original through unchanged, which is handy for A/B wiring. saved_path is the file on disk, and report_json carries the audit.

That audit is the reason to trust this node at all. The interpolation worker runs isolated, under a serial GPU lease and an NVML resource guard that aborts if output disk free drops below 256 MB, then compares audio packet counts and timestamps and decoded PCM digests between input and output, plus the video timeline. Audio is copied, not re-encoded. Once the progress bar passes roughly 90%, you're watching media auditing, not inference - it hasn't hung.

Install and gotchas

Manager → MiniMax H3 Audio T8 → install → full restart. The pack's requirements.txt is intentionally empty, so this adds no Python packages and can't disturb your Torch stack.

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Do not attach SaveVideo afterwards. The node already writes the deliverable, and the pack's release notes describe exactly why that rule exists: an earlier default stored every frame losslessly, turned a 15-second clip into a 12.9 GB file and filled a drive.

The other trap is mixing jobs. Upscale and interpolate are separate stages with separate models - use the pack's Upscale_Then_Interpolation workflow, which runs them one after the other, instead of piling filters into one node. And no, it will not touch H3's sampling: this runs after your clip exists, and hand-testing has only produced accepted samples for the interpolated shorts the author reviewed, not a guarantee for fast motion or HUD-heavy footage.