Add MiniMax H3 DMAD full-critic 4-step LoRA
Browse files
loras/minimax_h3_dmad_full_critic_4step_NOTICE.txt
ADDED
|
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
DMAD Full Critic 4-Step MiniMax-H3 LoRA
|
| 2 |
+
|
| 3 |
+
Original authors: Zhengming Yu et al. (Texas A&M University and ByteDance).
|
| 4 |
+
Project: https://github.com/Yzmblog/DMAD
|
| 5 |
+
Paper: https://arxiv.org/abs/2610.02188
|
| 6 |
+
Source: https://huggingface.co/ZhengmingYu/DMAD/blob/4e61167e7d51573c889ac78a4f66e5085757af95/minimax_h3/dmad_minimax_h3_4step_full_critic.safetensors
|
| 7 |
+
SHA256: ac23eaef3a359666610769e4ef5f8af9530262c265f25ac211b0b54e51efc91e
|
| 8 |
+
|
| 9 |
+
The safetensors file is byte-identical to the upstream release; only its filename changed.
|
| 10 |
+
624 BF16 tensors, rank 128, alpha 128, adapter strength 1.0.
|
| 11 |
+
The full_critic name describes the training critic, not a separate inference model.
|
| 12 |
+
|
| 13 |
+
License: MiniMax H3 Community License Agreement (Model Derivative of MiniMax-H3).
|
| 14 |
+
The upstream authors modified MiniMax-H3 by LoRA fine-tuning.
|
| 15 |
+
Base model: https://huggingface.co/MiniMaxAI/MiniMax-H3
|
| 16 |
+
License: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
|
| 17 |
+
|
| 18 |
+
Recommended inference: DMAD re-noise sampler, 4 steps, video shift 12, audio shift 2,
|
| 19 |
+
CFG 1, 1344x768, 124 frames at 24 fps, stereo audio. Eight steps can improve fast motion.
|
| 20 |
+
WanGP accelerator: DMAD Full Critic 4 Steps, for MiniMax H3 FL2VA Pruned.
|
loras/minimax_h3_dmad_full_critic_4step_bf16.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ac23eaef3a359666610769e4ef5f8af9530262c265f25ac211b0b54e51efc91e
|
| 3 |
+
size 1383674320
|