沐阳 H3 · 二采放大设置

H3's first pass is fast and often a bit soft. The second pass is the fix - upscale, redraw at low denoise, keep the audio. 沐阳 H3 · 二采放大设置 (H3DetailSettings) is the standalone control panel for that second pass when you're using 沐阳 H3 · 长视频. It exists because a long-video node can't realistically hold twenty upscaling widgets in its own body without becoming unusable - so the settings live in their own node, and H3LongVideo exposes a single socket that takes the whole bundle.

The output is a single 二采设置 (MYANG_H3_DETAIL) value - a settings dict, not an image. Wire it into H3LongVideo's 二采设置 input, and the long-video loop applies it one segment at a time. Disable it and the long video just runs first-pass only.

The inputs that matter

• enabled - the master switch. Off by default; nothing else matters until this is on.
• mode - 放大 + 二采 (upscale then refine, the default and recommended), second pass at the same resolution, or upscale-only with no second pass.
• upscale_method - the big choice. neural_3d (neural 3D latent upscale, marked recommended), pixel/VAE, latent, or NVIDIA RTX VSR. The neural 3D path needs the LBH-123-AI H3 Latent Upscaler weights in ComfyUI/models/latent_upscale_models/ - download them separately.
• resolution / width / height - the output target; custom width/height only apply when you pick 自定义 for the output short edge.
• steps (4 recommended) and denoise (0.2 recommended, keep it 0.15–0.25) - the redraw contract. Low denoise = the upscale decides the look, the model just cleans it up.
• scheduler / sampler_name - beta / res_multistep defaults. Note this is deliberately not the Turbo contract - the second pass runs on the base model.
• latent_precision - fp16 (recommended, memory-cheap); switch to fp32 if you see odd color blocks.
• latent_chunk_steps - temporal chunking for the neural 3D upscaler: 8 to save memory, 16 default, 32 faster.
• passes - how many refine rounds; only the first pass upscales, later rounds refine at the same resolution.
• seed_mode - reuse the same seed every round (default) or not.
• 二采模型 (optional) - the Ref2VA base model, before the Turbo LoRA. Required whenever the second pass is enabled - the second pass must not run through a Turbo LoRA.

Install and ground rules

Pack install: ComfyUI Manager search "ComfyUI-MiniMaxH3-Myang", or git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang into custom_nodes, restart. No extra Python deps. The neural-3D weights and the RTX VSR path (separate NVIDIA VFX runtime) are the only extra downloads, and they're optional.

Final audio always comes from the first pass - the second pass only gets seam handling and duration trimming, which is a deliberate choice to keep the soundtrack untouched. Two practical notes: start the neural 3D path at fp16 / chunk 16, dropping the chunk to 8 if memory bites; and if the first pass is already at your target resolution, mode "second pass at same resolution" exists precisely so you're not upscaling just to redraw. And H3's weights themselves remain territory-restricted (US/EU/UK/South Korea excluded) - the pack can't fix that, so check before you commit.