MiniMax H3 Director Refine

H3's happy place is a small canvas. The Director's default first pass is 864×480 - 0.4MP, deliberately, because that's where a 33B video model actually stays coherent. The problem is nobody wants to watch 480p. The classic answer is "generate low, upscale after," and that's exactly what MiniMax H3 Director Refine exists to automate - except it does the upscale the way a video diffusion model wants it done, not the way an image model does.

This node is a config pack, not a sampler. It has no image output and never samples on its own. You attach it to the Director's refine socket, tell it what kind of second pass you want, and the Director runs each segment through it. Leave the socket empty and behavior is unchanged: one pass, done.

The three modes

The mode menu is where you pick your philosophy:

• refine - same-resolution second sample. No bigger canvas, just a fresh KSampler pass over the first result to clean it up. Great with a different, stronger UNET (refine_model) or after removing a first-pass Turbo LoRA.
• upscale - enlarge to a target canvas first, then second-sample at the bigger size. This is the classic generate-then-polish-at-resolution pipeline, but for video latents.
• latent_upscale - enlarge the H3 video latent only, no second sample. Fastest, least "re-painted", but you're relying on the upscaler model alone to add detail.

For upscale, the upscale_method decides how the first pass gets bigger. h3_latent upscales the H3 video latent with a 3D upscale model from models/latent_upscale_models/ - the option built for this model. lanczos is plain pixel interpolation, optionally fed through an upscale_model like RealESRGAN. nvidia_rtx_vsr uses NVIDIA's RTX Video Super Resolution, which needs the optional nvidia-vfx package and an NVIDIA GPU - the README is explicit that it's not a hard dependency because it would break AMD and cloud installs.

What you must wire

The part that trips everyone: refine and upscale modes require a sigmas input. This node doesn't generate its own noise schedule - you connect a ComfyUI BasicScheduler or ManualSigmas, pointing it at the same model the second pass will use (the Director's main model, or your refine_model if you're swapping UNETs). H3's sigma shift is applied internally by the Refine, so you don't add a SigmaShift node here. The example workflow in the pack shows it: BasicScheduler → Refine.sigmas.

The settings worth knowing

• passes - how many refine rounds. Default 1. Only the first pass enlarges; later passes are same-resolution refines. Yes, it goes to 9999. No, you don't need that.
• skip_fl2v - default on. Second-pass sampling changes the picture, and it loves to drift those pinned first/last keyframes. The node skips fl2v shots by default to protect them; turn it off if you want fl2v refined too.
• confirm_first_pass - a two-queue workflow. First Queue runs the first pass and writes a cache, blocking the main images output; second Queue with the same seed skips straight to refine. Useful when you want to eyeball the 480p result before spending the compute, but it demands a fixed seed.
• seed_mode - inherit reuses the Director's seed; offset walks it +1, +2 per pass.
• Canvas controls (aspect_ratio, megapixels, custom width/height) set the target after upscale - the Director stays at its own first-pass resolution. 1.0MP at 16:9 lands around 1376×768.

Outputs and troubleshooting

refine is the pack that plugs into the Director; width and height just report the resolved target canvas. After a run, the Director's images is the refined result and images_pre_refine is the untouched first pass - keep both if you want to judge whether the second pass earned its compute.

Where it goes wrong: forget the sigmas wire and refine silently won't run the way you expect; expect it. Upscaling from 0.4MP to real 720p/1080p will happily eat VRAM, so keep an eye on memory between segments. And if you're in a region excluded by the H3 license (US, EU, UK, South Korea), that restriction applies to the whole pipeline - the refine pass doesn't change what you're allowed to run.