| --- |
| tags: |
| - minimax-h3 |
| - vae |
| - image |
| - comfyui |
| --- |
| |
| # Experimental MiniMax H3 single-image VAE |
|
|
| An experimental image-specialized MiniMax H3 VAE that directly decodes a single temporal latent (`T=1`) into one image. It was made to provide a usable direct-one-latent image path, not as a claim of state-of-the-art or high-fidelity image reconstruction. It is distributed as a standard merged H3 VAE checkpoint: no custom node or separate decoder head is required. |
|
|
| > [!IMPORTANT] |
| > **This checkpoint is for images only. Do not use it as a replacement for the original H3 VAE in video workflows.** Its image-specialized decoder materially regresses multi-frame video reconstruction and can introduce patch-grid ghosting and cross-frame mixing. Keep the original MiniMax H3 VAE for video. |
|
|
| > [!NOTE] |
| > Reconstruction quality is limited. Outputs can remain soft and can lose fine text, thin contours, hair, foliage and microtexture; the example below is intended to show the actual result rather than imply a quality guarantee. |
|
|
| ## Example |
|
|
| Source images are shown on the left and reconstructions from this VAE on the right. |
|
|
|  |
|
|
| ## How it was trained |
|
|
| The original H3 encoder was frozen while the full decoder and `post_quant_conv` were fine-tuned to reconstruct the real source image directly from one H3 temporal latent. Training used a progressive 256→384 px curriculum and one pass over 51,083 unique images (41,259 Booru Essence images and 9,824 pseudo-camera photographs), with a decoder-aware objective combining Charbonnier reconstruction, edge, SSIM, and low-weight FDL/FDL-PIPS losses. The selected step-1597 decoder was merged back into the standard H3 VAE checkpoint, so it loads through normal H3 VAE tooling. |
|
|
| ## Validation |
|
|
| On a balanced unseen 512 px set containing 32 photographs and 32 art images, the selected checkpoint measured: |
|
|
| - **30.4425 dB PSNR** |
| - **0.939257 SSIM** |
| - **0.016817 MAE** |
|
|
| The two live ComfyUI examples above reached 31.8420 dB / 0.898509 SSIM and 32.0064 dB / 0.940569 SSIM. Additional deployment checks from approximately 1 to 3 megapixels reached 31.55–33.43 dB PSNR. |
|
|
| ## ComfyUI usage |
|
|
| Download `minimax_h3_t1_image_vae_step1597.safetensors` into: |
|
|
| ```text |
| ComfyUI/models/vae/ |
| ``` |
|
|
| Load it with ComfyUI's standard **Load VAE** node and use the normal VAE encode/decode nodes. The model expects MiniMax H3's 24-channel latent format and is intended for single-image (`T=1`) round trips. |
|
|
| ## Checkpoint |
|
|
| - Size: 5,207,808,784 bytes |
| - SHA-256: `6c3d0bfa055986a803a566a862fcde283a1e63db62829e5ef4a2a5aebf50bb86` |
| - Training checkpoint: step 1597 |
|
|
| Model usage remains subject to the applicable MiniMax H3 license and terms. |
|
|
