CineTimeline｜后处理模式

One dropdown, and it's the label on a switch

CinePostprocessMode is a single widget: mode, with three choices, output as a CINE_POSTPROCESS_MODE. It does no work on pixels whatsoever. It exists so that "which finishing path does this shot take" is a value you can set in one place - including from an API prompt - instead of a wiring change you make by hand in the graph.

Older versions of this pack baked the enhancer choice into the node that ran the enhancement (four tiers: off, an LTX detailer fast mode, an LTX fidelity mode, and a SeedVR2 7B final). Same idea, worse ergonomics: to A/B two options you had to reconnect links. Splitting the choice from the execution is what these two nodes do, and this is the choice half. On its own, disconnected from the selector, it's inert.

The three modes, and what each is actually good for

原生输出（不放大） - no upscale, no second sampling pass. In the pack's architecture this isn't "worse quality for free": because the branches are lazily routed, the upscale nodes, the VAE re-encode and the refinement sampler are skipped entirely, so a segment costs exactly one first pass. That's the right default for previewing a timeline, iterating on prompts, or shots where the first pass already looks right. New graphs ship defaulting to this.

FlashVSR x2 - the open video upscaler that broke through in October 2025 and got genuinely widely adopted for speed rather than quality. The mid-2026 community read is that it's fading while SeedVR2 takes the heavy-lift work; the useful split is FlashVSR when the source is already decent and you want it fast, something with more temporal awareness when the source is genuinely low-res. On a 720p-ish H3 render going up 2×, that's exactly the "already decent" case.

latent x2 H3精修 - H3's own refinement: VAE re-encode plus a low-denoise second sampling pass on the video latent. The important design detail is audio. The experimental second-pass in this pack takes the final AV latent and forces the first pass's audio latent back in, so the refine pass improves picture and cannot regenerate dialogue. If you've ever had a "quality pass" give you a different take of the same line, you know why that's the correct call.

That mapping is worth internalising because it decides the audio you hear: the mode chooses a video route, and the audio that comes back with it is the audio belonging to that route.

Legacy values still resolve

select() normalises the older H3精修 spelling to latent x2 H3精修, and VALIDATE_INPUTS still accepts a legacy RTXVSR x2 value even though it's no longer offered in the dropdown. That's compatibility, not a suggestion - old saved graphs keep working and keep their original wiring, and you shouldn't hand-edit a workflow to reintroduce it. Anything genuinely unrecognised raises Unknown postprocess mode.

Traps

The classic complaint - "I switched the mode and nothing changed" - is almost always a missing connection on the other node, not a bug here. This node names the branch; CinePostprocessSelector is what actually routes frames and audio.

Also don't assume the mode implies its dependencies exist. The refine chains are built from other people's packs - an RTX-flavoured experimental path wants Nvidia_RTX_Nodes_ComfyUI and ComfyUI-PT_H3ConcatAVLatent, and FlashVSR wants a FlashVSR node pack and its model weights. Nothing about choosing a mode installs anything for you.

Installing it

Part of ComfyUI-CineTimeline. Manager search CineTimeline, or:

cd ComfyUI/custom_nodes
git clone https://github.com/bo341805sg/ComfyUI-CineTimeline
# restart ComfyUI, hard-refresh the browser

No models, no pip dependencies from the pack itself. If you're on the INT8 ConvRot H3 baseline, remember the launcher flag - --disable-dynamic-vram - because that failure happens long before anything postprocess-related ever runs.