MiniMax H3 一键高清放大 - Star7

Sampling MiniMax H3 at 4K for ten seconds isn't a thing most cards can do, and sampling at 1MP and pretending it's 4K isn't either. The old image-gen answer to this was hi-res fix: generate at native resolution, upscale the latent, then run a few low-denoise steps to let the model re-commit detail at the new size. MiniMaxH3OneClickHDStar7 is that recipe, packaged for H3's audio-video latent, in one node.

What's inside the one click

Two stages, in this order.

A learned 3D latent upscale. The video latent gets normalised with the upscaler's own fixed 24-channel mean/std, pushed through a 3D latent upscaler checkpoint, then de-normalised. That's a real learned spatial-temporal resizer, not bicubic on a tensor - which matters, because H3's latent is where the motion and the detail both live.

A short low-noise refine. The upscaled latent gets a noise_mask of ones over video and zeroes over audio, and then goes through SamplerCustomAdvanced with simple sigmas that only cover the tail of the schedule - that's what refine_strength controls. The default preset is 2 steps at 0.25. Then it checks the result is finite and reassembles the nested latent with the original audio member. The audio is never resampled, never re-diffused, never touched. For a model whose whole selling point is jointly-generated dialogue and room tone, that's the correct decision, and it's the difference between "sharper" and "subtly out of sync."

The output is a latent, not pixels. VAE decode stays outside, so the intended chain is:

Sampler "latent" -> One-Click HD -> MiniMaxH3ChunkedDecodeStar7 -> video combine
All-in-one Conditioning "context" ----^

Inputs that matter

enable_hd is the master switch, and turning it off returns your latent untouched - no model load, no second pass, no downloads. That's your A/B control.

h3_context is required and comes from the pack's MiniMax H3 多合一条件载入 - Star7 (All-in-one Conditioning). Without a model and a positive in it, the node raises an "HD context is incomplete" error. It wants that context because it needs the model, the prompt, and - usefully - any first/last keyframes you're generating from, which it re-encodes with the VAE at the new resolution so you're not stamping low-res endpoint latents onto a high-res pass.

target_megapixels (0.2–36) is the real size knob, aspect preserved. If the target is at or below your source size the node returns the latent unchanged and says so - it will not shrink anything.

preset is a paired steps-and-strength recipe, and the smart part is that it's profile-aware. Turbo-style checkpoints get 2–3 steps at 0.18–0.30; base or plain-LoRA H3 gets 4–6 steps, because it isn't distilled and shouldn't be treated as a few-step model. It detects Turbo/PDD by reading upstream names and mounted patches. With a PDD model it uses the trained partition tail and Euler; if a PDD file was loaded with a normal LoRA loader it errors instead of producing mush. The status line shows the exact sigma ladder used, from the actual model - no fixed shift guesswork, and sigmas minus one is your step count.

refine_steps / refine_strength / seed are what "自定义" (custom) unlocks when the presets aren't it.

upscale_model defaults to minimax_h3_latent_upscaler_3d_fp16.safetensors and downloads itself from the HF mirror, then Hugging Face, SHA-256 verified into ComfyUI/models/latent_upscale_models. Only that one file is exposed, deliberately - so the node can't silently grab an LTX upscaler or a BF16/FP32 sibling out of that shared folder and call it a different quality tier.

Tiling, and the attention override

enable_tiling is a pure VRAM lever for the refine pass. It splits each high-res prediction into an aspect-aware grid of overlapping 2D tiles - 16:9 asked for 16 tiles becomes 3×6 (18), square-ish gets 4×4 - with boundaries snapped to H3's 2×2 latent patches and cosine-weighted blending across tile_overlap. tile_count is treated as a floor, not an exact number. Tiling cuts peak refine VRAM and costs you model calls; it doesn't change sigmas, sampler, or audio. It has nothing to do with the decode node's internal VAE tiling.

second_pass_attention defaults to 继承一采 - inherit the first pass. The rest of the list is the pack's CK/SLA/Sol/Hybrid menu if you want the refine pass on a different kernel. Note it restores the pack's global attention config after the refine finishes, so an override here won't leak into a cached chunk node on your next queued run.

Install and where it bites

Same pack, one install: ComfyUI Manager → MiniMax H3 Activation Chunk - Star7, or

cd ComfyUI/custom_nodes
git clone https://github.com/star7code/minimax-h3-chunk-star7.git

Restart, then put the upscaler in models/latent_upscale_models or let the node fetch it. Pack dependencies (scipy, scenedetect, ultralytics) come along whether or not this node needs them.

The failures you'll actually meet: a missing or incomplete context wire; a non-nested latent handed in (it wants [B,24,T,H,W] video plus audio, and says so); an upscaler that fails its SHA check, with an error naming the directory and every source that failed; and time. Tiles and refine steps both cost real GPU minutes on a 33B video model - change one, then read the report string. For an A/B you don't need to delete the node; flip enable_hd off and compare.