| --- |
| license: other |
| license_name: minimax-h3-community-license-agreement |
| license_link: LICENSE |
| base_model: MiniMaxAI/MiniMax-H3 |
| library_name: comfyui |
| pipeline_tag: image-text-to-video |
| tags: |
| - minimax-h3 |
| - comfyui |
| - quantization |
| - int8 |
| - w4 |
| - nvfp4 |
| - video |
| - audio |
| - fl2va |
| - ref2va |
| --- |
| |
| # MiniMax-H3 Stock ComfyUI Quants |
|
|
| Community FL2VA and Ref2VA diffusion-transformer checkpoints for |
| [`MiniMaxAI/MiniMax-H3`](https://huggingface.co/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](https://huggingface.co/DmitryDB/MiniMax-H3-DynTime-sQKV) |
| repository. |
|
|
| ## Choose a checkpoint |
|
|
| Download one FL2VA or Ref2VA checkpoint from the same profile row. |
|
|
| | Profile | Direct downloads | File size | GPU class and quant layout | |
| |---|---|---:|---| |
| | **INT8 ConvRot HQ** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot-HQ.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot-HQ.safetensors?download=true) | 21.908 GiB | **32 GB+ · RTX 30/40.** 145 INT8 ConvRot + 63 BF16 semantic matrices. Largest BF16 island. A 24 GB RTX 4090 loader test offloaded about 0.955 GiB. | |
| | **INT8 ConvRot** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot.safetensors?download=true) | 20.940 GiB | **24 GB · RTX 30/40.** 170 INT8 ConvRot + 38 BF16 semantic matrices. Fully resident in the RTX 4090 loader test. | |
| | **INT8 ConvRot Lite** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot-Lite.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot-Lite.safetensors?download=true) | 20.330 GiB | **24 GB · RTX 30/40.** 185 INT8 ConvRot + 23 BF16 semantic matrices. Leaves more memory for the rest of the workflow. | |
| | **W8/W4 ConvRot** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-W8W4-ConvRot.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-W8W4-ConvRot.safetensors?download=true) | 13.565 GiB | **16 GB · RTX 30/40.** 86 W8 + 114 W4 main matrices; the eight token-refiner matrices remain BF16. | |
| | **W4 ConvRot** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-W4-ConvRot.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-W4-ConvRot.safetensors?download=true) | 10.067 GiB | **12 GB · RTX 30/40.** 200 W4 main matrices + 8 INT8 token-refiner matrices. | |
| | **W4 ConvRot Offload** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-W4-ConvRot-Offload.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-W4-ConvRot-Offload.safetensors?download=true) | 9.708 GiB | **8 GB + CPU offload · RTX 30/40.** All 208 main and token-refiner matrices use W4. | |
| | **NVFP4 HQ** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-NVFP4-HQ.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-NVFP4-HQ.safetensors?download=true) | 13.597 GiB | **16–24 GB · RTX 50/Blackwell.** 170 NVFP4 + 30 BF16 main matrices; the eight token-refiner matrices remain BF16. | |
| | **NVFP4** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-NVFP4.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-NVFP4.safetensors?download=true) | 10.862 GiB | **8–12 GB · RTX 50/Blackwell.** All 208 main and token-refiner matrices use block-scaled NVFP4. | |
|
|
| 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. |
|
|
| | Checkpoint | Loaded weights | Peak reserved | Free after load | Result | |
| |---|---:|---:|---:|---| |
| | `MiniMax-H3_*VA-INT8-ConvRot-Lite.safetensors` | 100% | 20.424 GiB | 2.072 GiB | PASS | |
| | `MiniMax-H3_*VA-INT8-ConvRot.safetensors` | 100% | 21.025 GiB | 1.471 GiB | PASS | |
| | `MiniMax-H3_*VA-INT8-ConvRot-HQ.safetensors` | 95.6% | 21.002 GiB | 1.494 GiB | PASS; about 0.955 GiB offloaded | |
|
|
| These are loader/kernel measurements, not complete prompt-to-decoded-video |
| VRAM peaks. |
|
|
| ## Quantization and preserved components |
|
|
| All 16 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, W8, and W4 weights use ConvRot/Hadamard rotation with group size 256, |
| per-row FP32 scales, and deterministic scale search. |
|
|
| | INT8 profile | BF16 attention-output blocks | BF16 MLP `fc2` blocks | Token refiner | |
| |---|---|---|---| |
| | `INT8-ConvRot-Lite` | 0, 1, 2, 3, 4, 5, 6, 7, 9, 15, 19, 38, 45, 49 | 49 | Eight BF16 matrices | |
| | `INT8-ConvRot` | 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 17, 19, 20, 27, 38, 43, 44, 45, 46, 47, 49 | 39, 45, 49 | Eight BF16 matrices | |
| | `INT8-ConvRot-HQ` | all blocks 0–49 | 29, 39, 44, 45, 49 | Eight BF16 matrices | |
|
|
| ## Stock layout versus DT-sQKV |
|
|
| | Feature | This repository | DT-sQKV repository | |
| |---|---|---| |
| | Attention storage | Fused `qkv_proj` | Separate `q_proj`, `k_proj`, `v_proj` | |
| | Attention execution | One fused call | Three projection calls | |
| | Original FP32 `time_embedder` | Replaced by measured time table | Retained and evaluated at runtime | |
| | `adaln_t_table` | FP32 `[4097,16]` | Absent | |
| | Per-block AdaLN | 51 independent FP32 rank-16 projections | 51 independent FP32 rank-16 projections | |
| | ComfyUI | Stock | Core patch required | |
|
|
| 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: |
|
|
| ```text |
| ComfyUI/models/diffusion_models/ |
| ``` |
|
|
| A complete workflow also needs the separately maintained Qwen3-VL MiniMax-H3 |
| text encoder and these shared VAEs: |
|
|
| | File | Role | |
| |---|---| |
| | `vae/MiniMax-H3_VideoVAE-FP16.safetensors` | Video latent encoder and decoder | |
| | `vae/MiniMax-H3_AudioVAE-FP32.safetensors` | Audio latent encoder and decoder | |
|
|
| 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. |
|
|
