MiniMax H3 Ref2VA — pruned BF16
Modified MiniMax H3 checkpoint for ComfyUI. This compresses the timestep-conditioning AdaLN projections while preserving the rest of the original transformer. No INT8, FP8, FP4, or other weight quantization is applied. This is an inference checkpoint, not a drop-in replacement for training the full-width original model.
The checkpoint is 40,225,724,440 bytes, versus 66,280,486,840 bytes for our lossless ComfyUI conversion of the original. The 51 AdaLN projections are reduced from 2,688 inputs to 8. Their weights and biases are BF16; the shared 1,025 × 8 interpolation table is FP32, as required by the ComfyUI implementation. Other tensors preserve their source dtype. ComfyUI evaluates the reduced projections in FP32 at runtime.
This differs from Comfy-Org's pruned_bf16 reference, whose reduced AdaLN projections are stored in FP16. BF16 has a wider exponent range but fewer mantissa bits, so this export has greater projection rounding error. It is not advertised as numerically identical or superior.
Sources and construction
- Original:
MiniMaxAI/MiniMax-H3, revision42ed227ee7df40d41602854ae760620d6eb651fe,transformer_ref/. - Coordinate table and comparison checkpoint:
Comfy-Org/MiniMax-H3, revisione5eb578a89295337b8ff433a035929ce0279e0b6. - Validated ComfyUI source:
e377e263049f9338b4d12a3dd417b36ae62948ff.
The conversion fits an affine map from the reference table to the original activated timestep embeddings, folds that map into each original projection, and rounds the resulting weights and biases to BF16. The published conversion script records the procedure; it requires the original checkpoint and the reference coordinate table at its documented local paths.
All 429 non-AdaLN tensors match the Comfy-Org reference exactly after lossless key, QKV, and SwiGLU layout conversion. Every row in all 51 reduced projections was checked on 297 unique timestep probes including endpoints, grid points, off-grid points, and seeded random points. Maximum per-projection relative L2 error was 0.171% against the original projection evaluated in FP32. This is a numerical probe, not a general quality guarantee. Exact measurements are in validation/pruning-validation.json.
Loading
Place the safetensors file under ComfyUI/models/diffusion_models/, select it in the Load Diffusion Model node, and use MiniMax H3 Reference to Video conditioning with the H3 text encoder and video/audio VAEs. Those supporting weights are not included here. Use an H3-compatible ComfyUI version that supports adaln_t_table checkpoints. Match ordered <Picture N>, <Video N>, and <Audio N> references in the prompt to the conditioning inputs.
Generation comparison
All three checkpoints completed a matched Ref2VA generation: 1344 × 768, 124 native frames at 24 fps, seed 0, 30 steps, res_multistep sampler, beta scheduler, PyTorch SDPA, no LoRA. All used the same cached conditioning from the same BF16 text encoder and FP32 video/audio VAEs. Initial noise and sigma schedules were verified identical. The presentation copies retain the first 120 frames and matching audio (exactly five seconds). All have stereo audio; no non-finite outputs or clipped audio samples were found.
The prompt and input media URLs come from MiniMax's official five-second Ref2VA example. Exact configuration and source hashes are recorded in validation/generation-config.json.
| Checkpoint | Sampling seconds, including model transfer | Peak allocated GPU memory (GiB) |
|---|---|---|
| Original BF16 | 1133.7 | 72.13 |
| Our pruned BF16 | 1133.2 | 47.94 |
| Comfy-Org pruned reference | 1133.2 | 47.94 |
These are single-run measurements on an RTX PRO 6000 Blackwell, not a rigorous speed benchmark. The clear benefit here is weight storage and memory reduction; sampling time was essentially unchanged.
Decoded presentation-video pixel PSNR against the original was 41.84 dB for our pruned BF16 and 41.64 dB for Comfy-Org's reference. Waveform relative L2 differences were 0.2010 and 0.0889, respectively. These are divergence measurements, not perceptual quality scores. One prompt and seed cannot establish general quality equivalence; assess the examples and full results in validation/comparison.json. Original full-width AdaLN uses BF16 runtime arithmetic, while reduced AdaLN uses ComfyUI's FP32 path.
AI-generated comparison clips:
License and attribution
Powered by MiniMax H3. The MiniMax H3 Community License Agreement applies, including its territory and use restrictions; see LICENSE and NOTICE. This is an independently modified checkpoint, not an official MiniMax release. Any included generated examples are AI-generated.