MiniMax-H3 Stock ComfyUI Quants

Community FL2VA and Ref2VA diffusion-transformer checkpoints for MiniMaxAI/MiniMax-H3. All files in this repository retain all 50 transformer blocks and use the stock ComfyUI fused-QKV and time-table layout. No custom node or ComfyUI core patch is required.

These are community conversions, not official MiniMax or ComfyOrg releases.

Naming

• No runtime marker in the filename means stock ComfyUI compatible.
• FL2VA is text/first-frame/last-frame-to-audio-video generation.
• Ref2VA is reference-image/video/audio-to-audio-video generation.
• Quantized tensor counts, retained BF16 islands, GPU class, and expected memory class are documented here instead of being encoded in filenames.
• The patch-required dynamic-time, separate-QKV editions use the explicit DT-sQKV marker and live in the separate MiniMax-H3-DynTime-sQKV repository.

Choose a checkpoint

Download one FL2VA or Ref2VA checkpoint from the same profile row.

Checkpoint size is not full-workflow peak VRAM. Resolution, frame count, attention backend, text encoder, VAEs, and ComfyUI offload settings also affect memory use. RTX 50 classifications are architecture-based; no full generation run was performed on an RTX 5090. NVFP4 here is block-scaled NVFP4, not AWQ.

Measured RTX 4090 loader results

FL2VA and Ref2VA were tested independently through stock ComfyUI. Each test also executed a real quantized INT8 projection.

These are loader/kernel measurements, not complete prompt-to-decoded-video VRAM peaks.

Quantization and preserved components

All 10 diffusion checkpoints:

• retain all 50 transformer blocks;
• use the fused qkv_proj = cat(Q,K,V) layout expected by stock ComfyUI;
• use a rank-16 FP32, 4,097-point time table;
• retain 51 independent FP32 AdaLN projections;
• keep norms, conditioning projections, patch projections, output heads, and other small or sensitive tensors in source precision;
• load without a custom loader or core patch in tested ComfyUI commit 14b05228.

INT8 weights use ConvRot/Hadamard rotation with group size 256, per-row FP32 scales, and deterministic scale search.

Stock layout versus DT-sQKV

The time table does not remove timestep conditioning. It interpolates a compact representation of the original measured time curve. Maximum measured table interpolation error is below 0.001%; sampled end-to-end AdaLN relative error is approximately 3e-7 to 4e-7 across 19 timesteps.

Validation

Every released checkpoint passed:

1. exact key, shape, dtype, and quantization-inventory checks;
2. sampled reconstruction against its original FL2VA or Ref2VA HF shards;
3. a 19-timestep FP32 AdaLN numerical comparison;
4. complete CPU load as MiniMaxH3Model in clean ComfyUI commit 14b05228;
5. remote byte-size and LFS SHA-256 verification.

Reports under reports/ retain their historical internal profile names so the published validation provenance remains intact. BF16 samples were checked bit-for-bit. A representative INT8 QKV sample has relative L2 error 0.008814. A prompt-to-decoded-video perceptual A/B score has not been measured.

Installation and required components

Place one selected FL2VA or Ref2VA checkpoint in:

ComfyUI/models/diffusion_models/

A complete workflow also needs the separately maintained Qwen3-VL MiniMax-H3 text encoder and these shared VAEs:

No text encoder is included in this repository.

License and attribution

Use is subject to the included MiniMax-H3 community license. The base model is by MiniMax. This community conversion is not endorsed by MiniMax or ComfyOrg.