Hugging Face's logo

DmitryDB
/
MiniMax-H3-ComfyUI-Quants

Image-Text-to-Video
MiniMax H3
comfyui
quantization
int8
nvfp4
video
audio
fl2va
ref2va
Model card Files
xet
 Community
2
DmitryDB commited on
Commit
adc7972
·
verified ·
1 Parent(s): 9543d2e

Update model card, filenames, and checksums

Browse files

Document GPU memory classes, quant layouts, stock versus DT-sQKV compatibility, and the renamed public artifacts.

Files changed (4) hide show
  1. FL2VA/SHA256SUMS +8 -8
  2. README.md +61 -75
  3. Ref2VA/SHA256SUMS +8 -8
  4. vae/SHA256SUMS +2 -2
FL2VA/SHA256SUMS CHANGED
@@ -1,8 +1,8 @@
1
- ad0cf8b50ba92068b5a4477527d5da73412a22db99cded35b07195504e78cf6a minimax-h3-fl2va-int8-convrot-balanced.safetensors
2
- 035312b34614cdc391722dc0d7e9a77a69b2d568ae9079fe629c5efc6457d4a0 minimax-h3-fl2va-int8-convrot-max.safetensors
3
- 95833a86493b24bcc4d142f541cbc14a7b27de6c5e29a829ff65c40033d91db6 minimax-h3-fl2va-int8-convrot-safe.safetensors
4
- 1519c19084ba0387a5f5c3983fcf32aee5923045de1ef4e44c04ca2e2ba486e5 minimax-h3-fl2va-nvfp4-compact.safetensors
5
- 1646d1148c0c9a7c2f728c8dbcfe7167e24cc138031feebfd44e04b2daac161f minimax-h3-fl2va-nvfp4-quality.safetensors
6
- 73fd34baee103b27f7de5be31b42e8680406816d0eb293f3a9d592fc10947f4a minimax-h3-fl2va-w4-convrot-compact.safetensors
7
- a9d64ad6b7970a9fa7620c8b82cbbec0c8e475310cf6c4c064fcda5d5ee94efb minimax-h3-fl2va-w4-convrot-offload.safetensors
8
- 45aebc6f6ba56a73e32e0cdabb3f039c8346528a9bc4cfc842de6346afdcd658 minimax-h3-fl2va-w8w4-convrot-balanced.safetensors
 
1
+ 035312b34614cdc391722dc0d7e9a77a69b2d568ae9079fe629c5efc6457d4a0 MiniMax-H3_FL2VA-INT8-ConvRot-HQ.safetensors
2
+ 1519c19084ba0387a5f5c3983fcf32aee5923045de1ef4e44c04ca2e2ba486e5 MiniMax-H3_FL2VA-NVFP4.safetensors
3
+ 1646d1148c0c9a7c2f728c8dbcfe7167e24cc138031feebfd44e04b2daac161f MiniMax-H3_FL2VA-NVFP4-HQ.safetensors
4
+ 45aebc6f6ba56a73e32e0cdabb3f039c8346528a9bc4cfc842de6346afdcd658 MiniMax-H3_FL2VA-W8W4-ConvRot.safetensors
5
+ 73fd34baee103b27f7de5be31b42e8680406816d0eb293f3a9d592fc10947f4a MiniMax-H3_FL2VA-W4-ConvRot.safetensors
6
+ 95833a86493b24bcc4d142f541cbc14a7b27de6c5e29a829ff65c40033d91db6 MiniMax-H3_FL2VA-INT8-ConvRot-Lite.safetensors
7
+ a9d64ad6b7970a9fa7620c8b82cbbec0c8e475310cf6c4c064fcda5d5ee94efb MiniMax-H3_FL2VA-W4-ConvRot-Offload.safetensors
8
+ ad0cf8b50ba92068b5a4477527d5da73412a22db99cded35b07195504e78cf6a MiniMax-H3_FL2VA-INT8-ConvRot.safetensors
README.md CHANGED
@@ -18,106 +18,92 @@ tags:
18
  - ref2va
19
  ---
20
 
21
- # MiniMax-H3 Quants for ComfyUI
22
 
23
- Community quantized diffusion-transformer checkpoints for
24
  [`MiniMaxAI/MiniMax-H3`](https://huggingface.co/MiniMaxAI/MiniMax-H3).
25
- This repository provides the same eight precision profiles for both **FL2VA**
26
- and **Ref2VA**. All files use the stock ComfyUI fused-QKV and time-table layout:
27
- no custom node or core patch is required.
28
 
29
  These are community conversions, not official MiniMax or ComfyOrg releases.
30
 
31
- ## 1. Choose FL2VA or Ref2VA
32
 
33
- - **FL2VA** — text-to-audio-video, optionally conditioned by a first frame,
34
- last frame, or both.
35
- - **Ref2VA** — reference-to-audio-video using reference images, video, and/or
36
- audio.
 
 
 
 
 
37
 
38
- Download one diffusion checkpoint from the matching column below.
39
 
40
- ## 2. Choose a quant
41
 
42
- | Profile | Direct downloads | Size | What it contains and when to use it |
43
  |---|---|---:|---|
44
- | **INT8 ConvRot balanced** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/minimax-h3-fl2va-int8-convrot-balanced.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/minimax-h3-ref2va-int8-convrot-balanced.safetensors?download=true) | 20.940 GiB | **Recommended for RTX 3090/4090 24 GiB.** 170 INT8 + 38 BF16 semantic matrices. Best tested quality/VRAM balance and fully resident in the RTX 4090 loader test. |
45
- | **INT8 ConvRot safe** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/minimax-h3-fl2va-int8-convrot-safe.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/minimax-h3-ref2va-int8-convrot-safe.safetensors?download=true) | 20.330 GiB | 185 INT8 + 23 BF16 semantic matrices. Choose this on a 24 GiB RTX 30/40 card when the rest of the workflow needs more VRAM. |
46
- | **INT8 ConvRot max** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/minimax-h3-fl2va-int8-convrot-max.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/minimax-h3-ref2va-int8-convrot-max.safetensors?download=true) | 21.908 GiB | 145 INT8 + 63 BF16 semantic matrices. Largest BF16 quality island; recommended for 32 GiB or more. About 0.955 GiB was offloaded in the 24 GiB RTX 4090 loader test. |
47
- | **W8/W4 ConvRot balanced** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/minimax-h3-fl2va-w8w4-convrot-balanced.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/minimax-h3-ref2va-w8w4-convrot-balanced.safetensors?download=true) | 13.565 GiB | 86 W8 + 114 W4 main matrices; BF16 token refiner. Starting point for 16 GiB RTX 30/40 cards. |
48
- | **W4 ConvRot compact** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/minimax-h3-fl2va-w4-convrot-compact.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/minimax-h3-ref2va-w4-convrot-compact.safetensors?download=true) | 10.067 GiB | 200 W4 main matrices + 8 INT8 token-refiner matrices. Recommended starting point for 12 GiB cards. |
49
- | **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 | All 208 main/refiner matrices use W4. Smallest portable profile; intended for 8 GiB cards with CPU offload. |
50
- | **NVFP4 quality** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/minimax-h3-fl2va-nvfp4-quality.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/minimax-h3-ref2va-nvfp4-quality.safetensors?download=true) | 13.597 GiB | 170 NVFP4 + 30 BF16 main matrices; BF16 token refiner. Recommended for RTX 50/Blackwell 16–24 GiB when NVFP4 support is available. |
51
- | **NVFP4 compact** | [FL2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/minimax-h3-fl2va-nvfp4-compact.safetensors?download=true) · [Ref2VA](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/minimax-h3-ref2va-nvfp4-compact.safetensors?download=true) | 10.862 GiB | All 208 main/refiner matrices use block-scaled NVFP4. Smallest Blackwell-specific profile for 8–12 GiB cards. |
52
-
53
- ### Short answer
54
-
55
- - **RTX 4090 24 GiB:** start with `int8-convrot-balanced`; use `safe` if the
56
- workflow needs more activation memory.
57
- - **RTX 5090 / Blackwell 16–24 GiB:** start with `nvfp4-quality` for headroom,
58
- or INT8 balanced when portability matters.
59
- - **RTX 30/40 16 GiB:** start with `w8w4-convrot-balanced`.
60
- - **12 GiB:** start with `w4-convrot-compact`.
61
- - **8 GiB:** use `w4-convrot-offload` and CPU offload.
62
- - **32 GiB or more:** `int8-convrot-max` has the largest BF16 island.
63
 
64
  Checkpoint size is not full-workflow peak VRAM. Resolution, frame count,
65
- attention backend, text encoder, VAE, and ComfyUI offload settings also matter.
66
- RTX 50 recommendations are architecture-based; no RTX 5090 generation run was
67
- performed on this machine. NVFP4 here is plain block-scaled NVFP4, not AWQ.
68
 
69
  ## Measured RTX 4090 loader results
70
 
71
  FL2VA and Ref2VA were tested independently through stock ComfyUI. Each test
72
  also executed a real quantized INT8 projection.
73
 
74
- | Profile | Loaded weights | Peak reserved | Free after load | Result |
75
  |---|---:|---:|---:|---|
76
- | `int8-convrot-safe` | 100% | 20.424 GiB | 2.072 GiB | PASS |
77
- | `int8-convrot-balanced` | 100% | 21.025 GiB | 1.471 GiB | PASS |
78
- | `int8-convrot-max` | 95.6% | 21.002 GiB | 1.494 GiB | PASS; about 0.955 GiB offloaded |
79
 
80
- These are loader/kernel measurements, not full prompt-to-decoded-video peaks.
 
81
 
82
- ## Compatibility
83
 
84
- All 16 files in this repository:
85
 
86
  - retain all 50 transformer blocks;
87
- - use fused `qkv_proj = cat(Q,K,V)` tensors expected by stock ComfyUI;
88
  - use a rank-16 FP32, 4,097-point time table;
89
  - retain 51 independent FP32 AdaLN projections;
90
- - load without a custom loader or core patch in the tested ComfyUI revision.
 
 
 
91
 
92
- For the original runtime FP32 time MLP and physically separate Q/K/V modules,
93
- use the patch-required
94
- [MiniMax-H3-DynTime-sQKV](https://huggingface.co/DmitryDB/MiniMax-H3-DynTime-sQKV)
95
- repository instead.
96
 
97
- <details>
98
- <summary><strong>Advanced: exact INT8 BF16 islands and time/QKV layout</strong></summary>
99
-
100
- INT8 weights use ConvRot/Hadamard rotation with group size 256, per-row FP32
101
- scales, and deterministic scale search. Norms, conditioning projections, patch
102
- projections, output heads, and other small or sensitive tensors retain their
103
- source precision.
104
-
105
- | Profile | BF16 attention-output blocks | BF16 MLP `fc2` blocks | Other main semantic matrices |
106
  |---|---|---|---|
107
- | `int8-convrot-safe` | 0, 1, 2, 3, 4, 5, 6, 7, 9, 15, 19, 38, 45, 49 | 49 | INT8 ConvRot |
108
- | `int8-convrot-balanced` | 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 | INT8 ConvRot |
109
- | `int8-convrot-max` | all blocks 0–49 | 29, 39, 44, 45, 49 | INT8 ConvRot |
110
 
111
- The eight token-refiner semantic matrices remain BF16 in all three INT8
112
- profiles.
113
 
114
- | Feature | This stock repository | Patched DynTime `s-QKV` repository |
115
  |---|---|---|
116
- | Attention | One fused projection call | Separate Q, K, and V calls |
117
- | Original FP32 `time_embedder` | Absent | Present |
 
118
  | `adaln_t_table` | FP32 `[4097,16]` | Absent |
119
- | `adaln_curve_basis` | Absent | FP32 `[2688,16]` |
120
- | `adaln_curve_mean` | Absent | FP32 `[2688]` |
121
  | Per-block AdaLN | 51 independent FP32 rank-16 projections | 51 independent FP32 rank-16 projections |
122
  | ComfyUI | Stock | Core patch required |
123
 
@@ -126,8 +112,6 @@ representation of the original measured time curve. Maximum measured table
126
  interpolation error is below `0.001%`; sampled end-to-end AdaLN relative error
127
  is approximately `3e-7` to `4e-7` across 19 timesteps.
128
 
129
- </details>
130
-
131
  ## Validation
132
 
133
  Every released checkpoint passed:
@@ -138,9 +122,11 @@ Every released checkpoint passed:
138
  4. complete CPU load as `MiniMaxH3Model` in clean ComfyUI commit `14b05228`;
139
  5. remote byte-size and LFS SHA-256 verification.
140
 
141
- Reports are stored under `reports/release_matrix/`. BF16 samples are checked
142
- bit-for-bit. A representative INT8 QKV sample has relative L2 error `0.008814`.
143
- A prompt-to-decoded-video perceptual A/B score has not yet been measured.
 
 
144
 
145
  ## Installation and required components
146
 
@@ -151,12 +137,12 @@ ComfyUI/models/diffusion_models/
151
  ```
152
 
153
  A complete workflow also needs the separately maintained Qwen3-VL MiniMax-H3
154
- text encoder and these shared VAE files:
155
 
156
  | File | Role |
157
  |---|---|
158
- | `vae/minimax_h3_video_vae_fp16.safetensors` | Video latent encode/decode |
159
- | `vae/minimax_h3_audio_vae_fp32.safetensors` | Audio latent encode/decode |
160
 
161
  No text encoder is included in this repository.
162
 
 
18
  - ref2va
19
  ---
20
 
21
+ # MiniMax-H3 Stock ComfyUI Quants
22
 
23
+ Community FL2VA and Ref2VA diffusion-transformer checkpoints for
24
  [`MiniMaxAI/MiniMax-H3`](https://huggingface.co/MiniMaxAI/MiniMax-H3).
25
+ All files in this repository retain all 50 transformer blocks and use the
26
+ stock ComfyUI fused-QKV and time-table layout. No custom node or ComfyUI core
27
+ patch is required.
28
 
29
  These are community conversions, not official MiniMax or ComfyOrg releases.
30
 
31
+ ## Naming
32
 
33
+ - No runtime marker in the filename means **stock ComfyUI compatible**.
34
+ - `FL2VA` is text/first-frame/last-frame-to-audio-video generation.
35
+ - `Ref2VA` is reference-image/video/audio-to-audio-video generation.
36
+ - Quantized tensor counts, retained BF16 islands, GPU class, and expected
37
+ memory class are documented here instead of being encoded in filenames.
38
+ - The patch-required dynamic-time, separate-QKV editions use the explicit
39
+ `DT-sQKV` marker and live in the separate
40
+ [MiniMax-H3-DynTime-sQKV](https://huggingface.co/DmitryDB/MiniMax-H3-DynTime-sQKV)
41
+ repository.
42
 
43
+ ## Choose a checkpoint
44
 
45
+ Download one FL2VA or Ref2VA checkpoint from the same profile row.
46
 
47
+ | Profile | Direct downloads | File size | GPU class and quant layout |
48
  |---|---|---:|---|
49
+ | **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. |
50
+ | **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. |
51
+ | **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. |
52
+ | **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. |
53
+ | **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. |
54
+ | **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. |
55
+ | **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. |
56
+ | **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. |
 
 
 
 
 
 
 
 
 
 
 
57
 
58
  Checkpoint size is not full-workflow peak VRAM. Resolution, frame count,
59
+ attention backend, text encoder, VAEs, and ComfyUI offload settings also affect
60
+ memory use. RTX 50 classifications are architecture-based; no full generation
61
+ run was performed on an RTX 5090. NVFP4 here is block-scaled NVFP4, not AWQ.
62
 
63
  ## Measured RTX 4090 loader results
64
 
65
  FL2VA and Ref2VA were tested independently through stock ComfyUI. Each test
66
  also executed a real quantized INT8 projection.
67
 
68
+ | Checkpoint | Loaded weights | Peak reserved | Free after load | Result |
69
  |---|---:|---:|---:|---|
70
+ | `MiniMax-H3_*VA-INT8-ConvRot-Lite.safetensors` | 100% | 20.424 GiB | 2.072 GiB | PASS |
71
+ | `MiniMax-H3_*VA-INT8-ConvRot.safetensors` | 100% | 21.025 GiB | 1.471 GiB | PASS |
72
+ | `MiniMax-H3_*VA-INT8-ConvRot-HQ.safetensors` | 95.6% | 21.002 GiB | 1.494 GiB | PASS; about 0.955 GiB offloaded |
73
 
74
+ These are loader/kernel measurements, not complete prompt-to-decoded-video
75
+ VRAM peaks.
76
 
77
+ ## Quantization and preserved components
78
 
79
+ All 16 diffusion checkpoints:
80
 
81
  - retain all 50 transformer blocks;
82
+ - use the fused `qkv_proj = cat(Q,K,V)` layout expected by stock ComfyUI;
83
  - use a rank-16 FP32, 4,097-point time table;
84
  - retain 51 independent FP32 AdaLN projections;
85
+ - keep norms, conditioning projections, patch projections, output heads, and
86
+ other small or sensitive tensors in source precision;
87
+ - load without a custom loader or core patch in tested ComfyUI commit
88
+ `14b05228`.
89
 
90
+ INT8, W8, and W4 weights use ConvRot/Hadamard rotation with group size 256,
91
+ per-row FP32 scales, and deterministic scale search.
 
 
92
 
93
+ | INT8 profile | BF16 attention-output blocks | BF16 MLP `fc2` blocks | Token refiner |
 
 
 
 
 
 
 
 
94
  |---|---|---|---|
95
+ | `INT8-ConvRot-Lite` | 0, 1, 2, 3, 4, 5, 6, 7, 9, 15, 19, 38, 45, 49 | 49 | Eight BF16 matrices |
96
+ | `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 |
97
+ | `INT8-ConvRot-HQ` | all blocks 0–49 | 29, 39, 44, 45, 49 | Eight BF16 matrices |
98
 
99
+ ## Stock layout versus DT-sQKV
 
100
 
101
+ | Feature | This repository | DT-sQKV repository |
102
  |---|---|---|
103
+ | Attention storage | Fused `qkv_proj` | Separate `q_proj`, `k_proj`, `v_proj` |
104
+ | Attention execution | One fused call | Three projection calls |
105
+ | Original FP32 `time_embedder` | Replaced by measured time table | Retained and evaluated at runtime |
106
  | `adaln_t_table` | FP32 `[4097,16]` | Absent |
 
 
107
  | Per-block AdaLN | 51 independent FP32 rank-16 projections | 51 independent FP32 rank-16 projections |
108
  | ComfyUI | Stock | Core patch required |
109
 
 
112
  interpolation error is below `0.001%`; sampled end-to-end AdaLN relative error
113
  is approximately `3e-7` to `4e-7` across 19 timesteps.
114
 
 
 
115
  ## Validation
116
 
117
  Every released checkpoint passed:
 
122
  4. complete CPU load as `MiniMaxH3Model` in clean ComfyUI commit `14b05228`;
123
  5. remote byte-size and LFS SHA-256 verification.
124
 
125
+ Reports under `reports/` retain their historical internal profile names so the
126
+ published validation provenance remains intact. BF16 samples were checked
127
+ bit-for-bit. A representative INT8 QKV sample has relative L2 error
128
+ `0.008814`. A prompt-to-decoded-video perceptual A/B score has not been
129
+ measured.
130
 
131
  ## Installation and required components
132
 
 
137
  ```
138
 
139
  A complete workflow also needs the separately maintained Qwen3-VL MiniMax-H3
140
+ text encoder and these shared VAEs:
141
 
142
  | File | Role |
143
  |---|---|
144
+ | `vae/MiniMax-H3_VideoVAE-FP16.safetensors` | Video latent encoder and decoder |
145
+ | `vae/MiniMax-H3_AudioVAE-FP32.safetensors` | Audio latent encoder and decoder |
146
 
147
  No text encoder is included in this repository.
148
 
Ref2VA/SHA256SUMS CHANGED
@@ -1,8 +1,8 @@
1
- f4b2a8b4a53446f5d696a39deb693b09ac7ce37013793a8b19340644e38b154f minimax-h3-ref2va-int8-convrot-balanced.safetensors
2
- 7693f972cea8b74aa9270550bee4c600cef88ea54f61fcb911ab138fde19b571 minimax-h3-ref2va-int8-convrot-max.safetensors
3
- 45ec67eaf82718c3f5a23bc3bb77b9d75f1d5a3a2436b6d5a17f183e47048327 minimax-h3-ref2va-int8-convrot-safe.safetensors
4
- f970a8a32ded2b4b72f68a624244f8e39d488b1fe0992b9de7133b400dbc95e2 minimax-h3-ref2va-nvfp4-compact.safetensors
5
- 10b4892c0aeb951666c0e8439ebc13d6360d2bfc52457244f37f9b09a3731ee3 minimax-h3-ref2va-nvfp4-quality.safetensors
6
- 870da69d24a26360c41947c842febbde69ee519428faa1d9a9aad5c4f7d3fa22 minimax-h3-ref2va-w4-convrot-compact.safetensors
7
- cc0763a63a1d6f7bb092337712b350fb96507f19be29b1d950de88140183d878 minimax-h3-ref2va-w4-convrot-offload.safetensors
8
- 11a9672f91c8b8a4b18234457f04f5969dcb522ae22a6c7eb8e5d2f472ad496d minimax-h3-ref2va-w8w4-convrot-balanced.safetensors
 
1
+ 10b4892c0aeb951666c0e8439ebc13d6360d2bfc52457244f37f9b09a3731ee3 MiniMax-H3_Ref2VA-NVFP4-HQ.safetensors
2
+ 11a9672f91c8b8a4b18234457f04f5969dcb522ae22a6c7eb8e5d2f472ad496d MiniMax-H3_Ref2VA-W8W4-ConvRot.safetensors
3
+ 45ec67eaf82718c3f5a23bc3bb77b9d75f1d5a3a2436b6d5a17f183e47048327 MiniMax-H3_Ref2VA-INT8-ConvRot-Lite.safetensors
4
+ 7693f972cea8b74aa9270550bee4c600cef88ea54f61fcb911ab138fde19b571 MiniMax-H3_Ref2VA-INT8-ConvRot-HQ.safetensors
5
+ 870da69d24a26360c41947c842febbde69ee519428faa1d9a9aad5c4f7d3fa22 MiniMax-H3_Ref2VA-W4-ConvRot.safetensors
6
+ cc0763a63a1d6f7bb092337712b350fb96507f19be29b1d950de88140183d878 MiniMax-H3_Ref2VA-W4-ConvRot-Offload.safetensors
7
+ f4b2a8b4a53446f5d696a39deb693b09ac7ce37013793a8b19340644e38b154f MiniMax-H3_Ref2VA-INT8-ConvRot.safetensors
8
+ f970a8a32ded2b4b72f68a624244f8e39d488b1fe0992b9de7133b400dbc95e2 MiniMax-H3_Ref2VA-NVFP4.safetensors
vae/SHA256SUMS CHANGED
@@ -1,2 +1,2 @@
1
- 7c1f131492e7eddacaac9069a61b81bdd39de5cc96561e677c5eab1cdce5e522 minimax_h3_video_vae_fp16.safetensors
2
- 8e505d95dd1561d47abd43d4238fd40d9bb1ae9e147ed0a4cba778d76ae4db48 minimax_h3_audio_vae_fp32.safetensors
 
1
+ 7c1f131492e7eddacaac9069a61b81bdd39de5cc96561e677c5eab1cdce5e522 MiniMax-H3_VideoVAE-FP16.safetensors
2
+ 8e505d95dd1561d47abd43d4238fd40d9bb1ae9e147ed0a4cba778d76ae4db48 MiniMax-H3_AudioVAE-FP32.safetensors