MiniMax H3 X2 Detail VAE: A 2-in-1 Upscale and Detail Release

MiniMax-H3-X2-Detail-VAE is a community release for MiniMax H3 that packs two things into one 5.2 GB checkpoint: a 2X video VAE that decodes H3 latents at double resolution, and an optional reference detail-enhancement node that rebuilds fine structure in an input image before reference-to-video. The author published the weights, a tested custom node and an example workflow on September 30, alongside a long writeup of why the original goal, a drop-in higher-detail VAE, was never reached.

MiniMax H3 X2 Detail VAE comparison frames

Frames from the release's comparison clip: plain 2X VAE decode on one side, the same latent with the detail-enhancement path on the other.

Two tools in one file

Most H3 VAE work on the wiki has been about making decode faster. This release goes the other way and asks whether decode can be made sharper, and it ships whatever survived that search. The same checkpoint, MiniMax-H3-X2-Detail-v1.safetensors, is used in two separate ways:

Mode	Input	What it does	Needs the custom node
2X VAE	H3 video latent	Decodes at twice the spatial resolution through a packed 12-channel output plus PixelShuffle	Yes, for the fast decode node only
Detail enhancer	RGB reference image	Rebuilds extra spatial structure from an early encoder tap before reference-to-video	Yes, included with the release

The two modes are not interchangeable. The 2X path works on a generated latent and needs no reference image. The detail path needs an existing RGB image, because the extra information it uses is pulled from the H3 encoder before the normal latent bottleneck.

Mode 1: 2X VAE decode

Put the checkpoint in ComfyUI/models/vae/ and select it in the standard Load VAE node. For the actual decode, replace ComfyUI's normal VAE Decode node with MiniMax H3 VAE Decode (fast) from TripleHeadedMonkey/ComfyUI-MiniMaxH3_LatentUpscaler, which is what turns the packed output into full-resolution RGB.

H3 video latent
   ↓
MiniMax H3 VAE Decode (fast)   ← MiniMax-H3-X2-Detail-v1
   ↓
packed 12-channel X2 decode
   ↓
PixelShuffle ×2
   ↓
2X RGB / video frames

The workflow ships with a validated starting configuration: tiling = true, tile_size = 256, tile_overlap = 64, output_device = cpu, temporal_tiling = false.

The MiniMax H3 VAE Decode (fast) node required by the 2X workflow

The required MiniMax H3 VAE Decode (fast) node, shown in the release's README.

Mode 2: reference detail enhancement

For image-reference workflows the same checkpoint is loaded a second time inside MiniMax H3 VAE 2X (Detailed Upscale), a small custom node bundled as ComfyUI-H3-X2-Detailed.zip. It sits between the source image and MiniMax H3 Reference to Video, and detail_strength = 1.0 is the tested baseline. Final video decode still runs through MiniMax H3 VAE Decode (fast) with the same X2 VAE.

The example workflow

The release includes a 2-in-1 graph that demonstrates both roles at once: the detail node enhances the RGB reference before reference-to-video, and the generated H3 latent is decoded through MiniMax H3 VAE Decode (fast) at the end.

The complete 2-in-1 workflow screenshot from the model card.

Head to head

Both of the clips below run the same model in its two released modes with matched generation settings: the left side is the 2X VAE on its own, the right side is 2X VAE plus the detailed path, where the reference is first processed through the B32 detail branch.

Comparison 01: 2X VAE decode versus 2X VAE with the detail-enhancement path.

Comparison 02: the same two modes on a second clip.

Why there is no "HQ VAE"

The writeup is unusually blunt about the result. The starting point was an already working H3 2X VAE whose larger output did not carry proportionally more real detail, so the project set out to make the extra pixels mean something. Decoder-side blocks, detail-target losses, latent "detail directions" and stronger progressive decoders all failed, and several of them improved a numeric loss while producing edge halos, engraving-like structure or nothing visible at all.

The one branch that clearly worked pulled a compact 32-channel representation from down.1.block.1 of the frozen H3 encoder, and it improved RMSE, HP3 and gradient metrics on all ten holdout frames. It also came from the original RGB image before the latent bottleneck, which is exactly the information that does not exist when H3 generates a fresh latent during text-to-video. The author's conclusion is narrow and stated plainly: the detail gain cannot be reproduced from the standard generated H3 latent alone, so it could not become the drop-in latent-only VAE the project was aiming for. What is left is an experimental 2-in-1 tool, not a solved bottleneck.

Requirements and files

MiniMax-H3-X2-Detail-v1.safetensors, the 5.2 GB checkpoint, goes in ComfyUI/models/vae/.
ComfyUI-MiniMaxH3_LatentUpscaler provides MiniMax H3 VAE Decode (fast) and is required for the 2X workflow.
ComfyUI-H3-X2-Detailed.zip adds the optional MiniMax H3 VAE 2X (Detailed Upscale) node for the reference path.
H3_VAE_Detailed_and_2X_VAE_2in1.json is the example workflow.

Availability

Weights and assets: speach1sdef178/MiniMax-H3-X2-Detail-VAE
ComfyUI file: MiniMax-H3-X2-Detail-v1.safetensors in models/vae/
Required node pack: TripleHeadedMonkey/ComfyUI-MiniMaxH3_LatentUpscaler
Base model: MiniMaxAI/MiniMax-H3