沐阳 H3 · 采样器（内部·步级预览）

During a long video generation you're watching the preview for feedback, and ComfyUI's default preview - the latent-to-RGB blurry smear - tells you almost nothing until the very end. 沐阳 H3 · 采样器（内部·步级预览） (H3SamplerAdvanced) is the pack's internal sampler that replaces that: at every sampling step it VAE-decodes a real frame and pushes it to the frontend. You watch the actual picture form, which for multi-minute long-video runs is the difference between "looks broken, cancel it" and "wait, it's actually coming together."

Functionally it's a drop-in for SamplerCustomAdvanced - same shape, same job - so it fits anywhere that node does. It takes the noise, guider, sampler, sigmas, and latent_image, runs the sampling loop, and returns the finished LATENT. The extra inputs (run_id, owner_id, segment_index, total_segments, pass_label) are the progress bookkeeping: they tag which run, which segment, and which pass each preview belongs to, so the Director's panel can file frames correctly. pass_label defaults to sample1 and becomes sample2 on a second pass - which is how the panel distinguishes first-pass from refined previews.

The preview itself is a proper decode: it pulls one temporal token from the middle of the current latent, decodes it with the VAE, and saves a PNG to ComfyUI's temp directory. That's more expensive per step than the blurry default, so it's a genuine trade - but for a workflow where a single segment takes minutes, the cost is noise and the visibility is the point.

Inputs

• noise / guider / sampler / sigmas / latent_image / vae - the standard advanced-sampler wiring.
• run_id / owner_id / segment_index / total_segments - progress identity.
• pass_label - which pass this sampling belongs to.

Output: output (LATENT), the sampled latents, exactly like the stock node.

Install and why you won't wire it

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.

Like the other 内部 nodes, the description says "请勿手动添加" - the Director and long-video expander place it for you. You'll meet it when you're curious why the previews are crisp, or when a custom graph needs the same step-level visibility. And the pack-wide reminder stands: H3's local weights are territory-restricted by community license (US/EU/UK/South Korea excluded), and after any ComfyUI update you should re-validate a two-segment run before trusting a long chain.