Minimax H3 Latent Upscaler (3D) — BF16 Conservative v5
MAXQUALITY re-quantization from FP32 source with conservative mixed-precision fallback.
📥 Download
| File | Size | Format |
|---|---|---|
minimax_h3_bf16_CONSERVATIVE_v5.safetensors |
~658 MB | BF16 + FP16 fallback |
🎯 What changed (vs original)
| Original BF16 (author) | This v5 | |
|---|---|---|
| Source | FP16 (naive conversion) | FP32 (full precision) |
| Strategy | .to(bfloat16) |
Conservative mixed-precision |
| Outlier correction | ❌ None | ❌ Not needed (clean BF16) |
| Mixed precision | ❌ 0% FP16 | ✅ 150 tensors FP16 (norm/head/tail/embed) |
| Conv layers | Distorted | Pure BF16 |
| Artifacts | Flickering, "soap" | None |
🚀 Usage (ComfyUI)
Place in: ComfyUI/models/latent_upscale_models/
Node settings:
precision:bf16device:cudatile_size: 256temporal_overlap: 2-4
🔬 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 pure: 172 tensors (all conv weights without modifications)
- Re-quantization: Conservative (no outlier correction, no smoothing, no calibration)
⚠️ License
This is a derivative work of Minimax H3 Latent Upscaler by LBH-123-AI.
License: Apache-2.0
You must comply with the original license terms. No additional restrictions applied.
🙏 Credits
- Original model: LBH-123-AI
- Based on: LTX 2.5 bf16 native approach