MiniMax-H3 VAE Loader / tile control (BFS)
The H3 VAE loader that exposes the tile size ComfyUI hides from you
- VAE
ComfyUI's native MiniMax H3 VAE loader has a quiet problem: the MiniMaxH3VideoVAE constructor hardcodes tile_size to 256, and the native loader doesn't expose the field. 256 happens to be the worst value you can measure - it multiplies seams. This node is the native loader plus the two knobs you actually want.
Use it when you already have a complete H3 VAE file - one with an encoder, e.g. an already-converted single-frame checkpoint. It loads it exactly like VAELoader does (same detection keys, same throw_exception_if_invalid), then lets you set:
vae_name- the full H3 VAE file, picked from yourComfyUI/models/vaefolder.tiling- default on. Keep it that way above ~768px; the decoder's training curriculum stopped at 1024, and whole-image decode above that shows a grid at 32px boundaries.tile_size- default 512, the measured optimum. On a 1056x640 image: PSNR 26.00 dB / seam 0.91 at 512, vs 22.17 / 1.49 at 256 (ComfyUI's default) and 21.35 / 2.42 at 1024 (above trained resolution). 475k of the 500k training images were ≤512px, so 512 keeps every tile inside the region with training mass.
Output is a single VAE, ready for the sampler or a decode node.
If you only have the decoder-only checkpoint
Don't use this node - it'll tell you exactly why. A decoder-only file lacks the encoder detection key, so ComfyUI would misdetect it as Stable Diffusion geometry and produce blocks; the error message points you at the other loader in this pack, MiniMaxH3SingleFrameVAELoader, which merges the decoder onto the official VAE for you.
That's also the answer to "why two loaders?" Different jobs: the single-frame loader builds a VAE from two files, this one loads an already-complete file and fixes the tiling ComfyUI won't let you touch. If you have a full file, this is the one you reach for.
Installing it
Same pack, same drill. ComfyUI Manager (search ComfyUI-BFSNodes) or:
cd ComfyUI/custom_nodes
git clone https://github.com/alisson-anjos/ComfyUI-BFSNodes.git
cd ComfyUI-BFSNodes
pip install -r requirements.txt
Restart, and put the full H3 VAE in ComfyUI/models/vae.
Common issues
The only realistic failure is feeding it a decoder-only file, and the node refuses loudly rather than producing garbage - that's the good kind of error. If you see "não será reconhecido como MiniMax-H3", that's it: either switch to the single-frame loader or give it the complete VAE. The rest of the time this node is boring in the best way - it's a stock loader with one default corrected.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| vae_name | COMBO | VAE completo do H3 (com encoder). Para o checkpoint decoder-only do autor, use o outro node. | |
| tiling | BOOLEAN | true | Tiling espacial. MANTENHA LIGADO acima de ~768px: o curriculo do autor parou em 1024 e a imagem inteira acima disso produz grade na fronteira de 32px. |
| tile_size | INT | 512128–2048 | 512 e o otimo MEDIDO: numa imagem 1056x640, PSNR 26.00 dB / costura 0.91, contra 22.17/1.49 com 256 (o default do ComfyUI) e 21.35/2.42 com 1024. O autor treinou 475k das 500k imagens em <=512px. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| VAE | VAE | — |