凤希AI - MiniMax放大器

Want a bigger MiniMax H3 video without re-running the sampler? That's what this node does: it takes the H3 latent that came out of sampling and makes it 2x, 3x, or 4x bigger before the VAE decodes it. Same video, higher resolution, no regeneration - and on a small card it stays cheap on VRAM, because it's one forward pass of a 3D network, not another diffusion run.

It's one node inside fxai-toolkit ("凤希全能节点包", aka FxAi), a big Chinese long-video pack whose whole thesis is that long AI video should run on 8–12GB cards: segment, loop, free VRAM, repeat. This upscaler is that same philosophy applied to output resolution.

Why upscale in latent space at all

The upscaling guide for this site beats this drum hard: upscaling video frame-by-frame is a trap, because nothing forces consecutive frames to agree. Decode → pixel-upscale → re-encode flickers on fine texture. This node sidesteps that by operating on the whole video latent at once, with temporal convolutions that see neighboring frames, so the upscale is coherent across time by construction. It's the "more pixels" job, not "more detail" - don't expect it to invent pores or fabric weave. It's cheap, temporally sane, and slots into exactly the spot a low-VRAM long-video pipeline wants: between the sampler and the VAE decode.

How it works

The node loads a dedicated upscaler network - a 3D-conv stack of res-blocks with temporal convs every couple of layers - and runs the latent through it. It normalizes the 24-channel H3 latent with hardcoded per-channel mean/std (H3-VAE-specific stats), resizes in the middle with trilinear interpolation, and passes a scale embedding into the network so it knows the target size. Output gets de-normalized and returned as a normal LATENT.

Two switches matter:

• 时间分块 (time chunking, default on) - processes the temporal axis in ~32-frame chunks with a 5-frame overlap and linear crossfade between chunks. This is the VRAM saver. On a 24GB card you can turn it off for a marginally cleaner result; on an 8GB card, leave it on.
• 释放显存 (VRAM release, default on) - after the run it moves the model back to CPU, soft-empties ComfyUI's cache and collects garbage. Free background behavior that keeps long workflows from slowly drowning.

It also auto-detects the architecture and dtype of whatever weights you drop in, and caches the loaded model between runs.

The inputs that actually matter

The widget labels are in Chinese - that's the pack's normal, don't let it throw you.

• 模型 (model) - a dropdown listing .safetensors/.pth files in models/latent_upscale_models. This is the one you must set up yourself (see install below). If it shows a placeholder starting with (请将模型放入... ("please put the model into..."), you haven't added the weights and the node will refuse to run.
• 放大倍数 (scale) - 2, 3, or 4, default 2. Start at 2; 4x on a 12GB card gets spicy.
• latent - the H3 video latent from your sampler.
• The only output is LATENT, which feeds straight into the VAE decode.

Installing it

Install the pack either through ComfyUI Manager (search "fxai-toolkit" or "凤希"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/fxai666/fxai-toolkit

Restart ComfyUI. There's no requirements.txt - the deps are torch, einops and safetensors, which ComfyUI already bundles, so no pip step. The step people actually miss is the model download: the tooltip on the model input credits LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler, which is where the weights come from. Grab the upscaler file (the pack's own workflows use minimax_h3_latent_upscaler_3d_bf16.safetensors), drop it into ComfyUI/models/latent_upscale_models/, and refresh the node.

Gotchas & troubleshooting

• The dropdown is empty or shows the placeholder - you haven't downloaded the model. That's the #1 "why won't it run" for this node, and the error is only in Chinese ("请将模型文件放入 latent_upscale_models 目录").
• Only feed it H3 latents. The normalization constants are baked in for the 24-channel H3 VAE. Wire it into an SDXL or Wan latent and you'll get noise-colored garbage, not a video. This is an H3-tool, full stop.
• The bf16 weights need a GPU that does bfloat16. The node detects and runs in whatever precision the file is. On a pre-Ampere card, look for a non-bf16 variant from LBH's repo rather than fighting it.
• Sometimes it silently does nothing. The target size is aligned to 32 pixels, and if the aligned result rounds back to the input size, the node just returns the latent unchanged. Output looks identical? That's why.
• License is murkier than it looks. The README says MIT, but the source files carry "商用需购买商业授权" (commercial use requires purchasing a license). If you're shipping content, resolve that with the author first.

Pack support is a Chinese QQ group and Bilibili channel (linked from the README); fine if you read Chinese, otherwise self-serve.

Need it in context? The pack's own MiniMax workflows show the full picture: split the audio-visual latent, upscale the video half here, stitch it back, decode. That's the pattern this node was built for.