MiniMax H3 Easy Segment Refine

Your first pass through a context-segment chain gives you connected shots at whatever resolution you set on the plan. If that's 480P or 540P - the safe place to start when you're iterating - you don't want to ship that. MiniMax H3 Easy Segment Refine is the pack's resolution-aware second pass: it re-samples each segment at a higher target while keeping the continuity, the per-shot prompts, and the first-pass audio intact. Same shots, more resolution, no scene-by-scene discontinuity.

What it does

This node takes the first-pass segments output from Segment Sample, the h3_context plan, a model, sampler, sigmas, and a seed, and runs each segment through a second sampling pass at a higher resolution. Crucially, it keeps per-segment boundaries honest: every segment rebuilds its own prompt and reference conditioning, and the refined tail - not the first-pass one - carries continuity into the next segment. The first-pass audio stream stays as the default timeline audio, so you're not re-rolling the soundtrack for free.

Picking a refine mode

The refine_mode input is the first decision:

• pixel_resize - decode each segment, resize it, re-encode. No extra model file required. The simplest and most predictable path, and it uses the first pass as the source.
• latent_upscale - instead of pixel resizing, upscale the latent with the H3 3D latent upscaler before re-sampling. This needs a matching upscaler checkpoint in models/latent_upscale_models/ - the node will refuse to run latent_upscale mode without one selected (latent_upscale_scale defaults to 1.3, and latent_upscale_device/latent_upscale_precision default to cuda/fp16).

Set your output size with target_width / target_height; leave them at 0 to follow the plan's canvas (the node treats 0 as "use the first pass's resolution").

The second decision is refine_execution. whole_segment re-samples each segment in one go. tiled_low_vram splits the current segment into spatial tiles (tile_width/tile_height default 512, tile_overlap 128, tile_fade 32) and samples them separately - it trades noticeably more sampling time for a much lower VRAM ceiling. If your card chokes on whole-segment refinement at 2K, tile it.

Where it sits

Context Segments → Segment Sample → Segment Refine → Segment Decode → final video

The output refined_segments (MINIMAX_H3_SEGMENTS) goes to Segment Decode, same as the first pass - you just swap which chain feeds it. The repo ships ready-made Pixel_Refine and Latent_Refine workflow files if you want the exact wiring without guessing.

Gotchas

Three things bite people. First, like Segment Sample, Segment Refine re-executes on every queue - no node-cache shortcuts, so tune your plan at low resolution before paying for a full refinement. Second, don't leave refine_mode on latent_upscale without the checkpoint installed; the validation explicitly blocks that combo with a clear message, so it's a fast fail rather than a silent wrong render. Third, if your refine pass and first pass disagree on continuity mode or context length, the chain's seams reappear - keep both nodes reading from the same Context Segments plan, and the whole thing stays coherent.

Installing

Same pack, same steps. Update ComfyUI to a build with the official MiniMax H3 nodes, then install ComfyUI-MiniMaxH3-Easy via ComfyUI Manager (search the pack, pick Nightly) or:

cd ComfyUI/custom_nodes
git clone https://github.com/nkxx188/ComfyUI-MiniMaxH3-Easy.git

Restart ComfyUI, and if you plan to use latent_upscale, drop the H3 upscaler checkpoint into models/latent_upscale_models/ before you queue anything.