沐阳 H3 · 二采放大精修（像素路径）

沐阳 H3 · 二采放大精修（像素路径） (H3DetailRefine) is the pixel-path member of the second-pass family. Where the long-video path hands everything to the 二采放大设置 node and picks upscaler methods like neural 3D latent, this node is the explicitly wired version for one particular recipe: upscale the first pass on the CPU, then redraw it with the H3 base model at low denoise. The "pixel path" in its name is the upscale method - plain pixel/VAE upscaling rather than latent tricks.

The mechanism is a clean three-step chain. First it validates you're not doing something contradictory: the model input must be a Ref2VA base checkpoint before any Turbo LoRA, and it rejects a Turbo model outright - Turbo's fixed NFE trajectory can't do a low-denoise beta-schedule redraw. It also checks the input frame count sits on H3's 17k+5 grid, because everything in this pack assumes the grid. Then it upscales to the target resolution in chunk_frames-sized groups (default 4) on the CPU, which keeps the big intermediate tensors out of VRAM. Finally it VAE-encodes the upscaled frames, re-encodes the audio if the sample rate doesn't match the audio VAE, and runs the low-denoise second pass - denoise defaults to 0.2 with beta scheduler and res_multistep.

The audio choice is worth knowing: the output is picture-refined, but the final audio for the delivered video remains the first pass's original soundtrack. The node takes audio as an input only so it can keep the timeline in sync - it doesn't regenerate sound.

Inputs

• h3 / model / conditioning - the loader bundle, the base Ref2VA model (pre-Turbo-LoRA), and the conditioning. model's tooltip is blunt: it must be the base model before the Turbo LoRA.
• images / audio - the first pass output you're refining.
• resolution / width / height - the target canvas (768P default, 1664×928).
• upscale_method / chunk_frames - pixel upscale by default; chunk size for the CPU pass.
• steps / denoise / scheduler / sampler_name / noise_seed - the redraw contract: 4 steps, 0.2 denoise, beta / res_multistep.

Outputs: refined_images (IMAGE) and detail_latent (LATENT) - the refined pictures and the latent they came from, if you want to keep going from latent space.

Install and the failure modes

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 two errors you'll actually hit are the two the node raises on purpose: wiring a Turbo-LoRA model in (it refuses loudly), and feeding it a frame count off the H3 grid (also refuses). Both are the pack protecting you from yourself. And the standing caveats: keep the second pass on a Ref2VA base without Turbo, and remember H3's weights are territory-restricted (US/EU/UK/South Korea excluded) - check the license before you build a project on them.