Minimax H3 Latent Upscaler (3D) — MAXQUALITY / NEXORA

This repository contains several checkpoints derived from the MiniMax H3 3D Latent Upscaler by LBH-123-AI.

It includes the original repository's BF16 conservative re-quantized checkpoint and two FP16 NEXORA checkpoints fine-tuned for latent upscaling on paired image and video latent data.

Important: The FP16 NEXORA checkpoints are fine-tuned versions of the existing upscaler. They are not models trained from scratch, and they are not LoRA adapters.

📥 Checkpoints

The two NEXORA files were saved as FP16. During training, model parameters were held in FP32, forward computation used FP16 autocast, the loss was calculated in FP32, and the final state dictionaries were converted to FP16 for saving.

🆕 NEXORA FP16 fine-tuning

What is NEXORA?

NEXORA is an experimental fine-tuning series built on the existing MiniMax H3 3D latent upscaler architecture. The objective was to adapt the upscaler using paired low-resolution guide latents and high-quality target latents from both images and videos.

The training objective was supervised latent-space fine-tuning:

1. Prepare corresponding guide and target media.
2. Encode the media into latent representations using the VAE workflow.
3. Feed the normalized low-resolution guide latent to the 3D upscaler.
4. Compare the predicted upscaled latent with the normalized high-quality target latent.
5. Calculate the training loss and update the model weights through backpropagation.

The training data was stored as prepared latent pairs, so the VAE encoding did not need to be repeated for every optimization step.

Training stages

The NEXORA checkpoints were produced in successive stages:

The 6,056-step Full-data stage is the final stage's step count. It is not 6,056 epochs. The 400-step checkpoint is an earlier checkpoint and was not overwritten by the Full-data run.

Full-data dataset

The final training run used the prepared full-data latent-pair dataset:

To avoid the much smaller video subset being overwhelmed by image examples, the final training schedule used each training image once and each training video twelve times:

• Images: 4,724 × 1
• Videos: 111 × 12
• Total optimization steps: 4,724 + 1,332 = 6,056

This corresponds to 6,056 optimization steps with an effective batch size of 1, not 6,056 epochs.

The prepared latent data used a 4× scale setting. The project history identifies 512 px as the preparation resolution; exact per-sample dimensions and media provenance are not specified in this README.

Training configuration

The recorded Full-data training configuration was:

The training loop checked that the loss and gradient norm remained finite. Gradients were unscaled before clipping, then the optimizer and GradScaler were stepped.

No EMA weights, LoRA training, or training-from-scratch stage are documented for these checkpoints. A learning-rate scheduler is not specified in the recorded final training configuration.

Validation results

The Full-data checkpoint was evaluated against the same held-out validation set before and after fine-tuning. The recorded relative changes were:

These are latent-space validation metrics (L1 and RMSE). They indicate that the predicted latents became closer to the target latents on the held-out samples.

These measurements are not a substitute for visual A/B testing. They do not by themselves prove that every image or video will look sharper, more detailed, or temporally more stable. Visual comparison in the intended ComfyUI workflow is still recommended.

Architecture

The NEXORA fine-tuning retained the existing LatentResizer3D architecture. The recorded configuration was:

The NEXORA checkpoints were validated against the expected model state structure; the project history records 322 tensors in the saved checkpoint. The architecture was not expanded with new layers for this fine-tuning run.

🎯 Original BF16 Conservative v5 checkpoint

The repository also retains the MAXQUALITY re-quantized checkpoint created from the FP32 source with a conservative mixed-precision fallback.

What changed vs. the original BF16 conversion

BF16 technical details

• Base model: LBH-123-AI/Minimax_h3_latent_Upscaler
• Source dtype: FP32 (minimax_h3_latent_upscaler_3d_fp32.pth)
• Target dtype: BF16 + FP16 fallback
• Total tensors: 322
• FP16 fallback: 150 tensors (GroupNorm, conv_in, conv_out, embed, emb_layers)
• BF16: 172 tensors
• Re-quantization: Conservative (no outlier correction, no smoothing, no calibration)

The BF16 v5 checkpoint and the NEXORA FP16 checkpoints are different model variants produced through different procedures. The NEXORA fine-tuning results should not be interpreted as a direct benchmark against BF16 v5 unless both are evaluated using the same visual and numerical test protocol.

🚀 Usage (ComfyUI)

For the original BF16 checkpoint, the documented location is:

ComfyUI/models/latent_upscale_models/

For the NEXORA FP16 checkpoints, use the same latent-upscaler model directory if your installed ComfyUI node supports the checkpoint architecture and FP16 weights.

Suggested starting settings for the original BF16 v5 checkpoint:

• precision: bf16
• device: cuda
• tile_size: 256
• temporal_overlap: 2–4

For NEXORA FP16 checkpoints, select FP16-compatible precision/loading settings in the relevant node. Exact node-specific settings may differ by ComfyUI implementation and GPU.

Always compare checkpoints with the same input, VAE, scale, tile settings, temporal overlap, and generation settings. Keep the source checkpoint unchanged when testing a new variant.

⚠️ License

This repository contains derivative model weights based on MiniMax H3 Latent Upscaler by LBH-123-AI.

The existing repository metadata declares Apache-2.0. Check the upstream model license and the licenses/terms of any training data or other incorporated materials before redistribution or commercial use. This README does not grant rights beyond the applicable upstream and third-party licenses.

🙏 Credits

• Original model architecture and checkpoint: LBH-123-AI
• NEXORA fine-tuning, experiments, and checkpoint preparation: Asirus
• Original BF16 v5 approach: based on the LTX 2.5 BF16 native approach, as described in the previous repository documentation.