Add MiniMax-H3-Image INT8 ConvRot checkpoints and VAE, update notice
Browse files
MiniMax-H3-Image-pruned_rank8_int8_convrot.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:118ddfce38dd29f0644f6ba374660b6fc83b6b68a2c468bd855332af4d6bb7d0
|
| 3 |
+
size 21066600232
|
MiniMax-H3-Image_int8_convrot.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6cec2c57a10a4f8be0848f1e9a04f43785fb62dbe587210e89e992ff3049b9e5
|
| 3 |
+
size 34047828466
|
minimax_h3_image/NOTICE
CHANGED
|
@@ -5,3 +5,5 @@ Source: https://huggingface.co/MiniMaxAI/MiniMax-H3-Image
|
|
| 5 |
The following files were modified by DeepBeepMeep for WanGP from the released MiniMax-H3-Image checkpoint:
|
| 6 |
- MiniMax-H3-Image_bf16.safetensors: the 12 transformer shards merged into one file; the per-step output heads of pdd_fused_heads.safetensors stored in it as pdd_heads.video_out.*, replacing the unused final_layer.video_out. All other tensors are unchanged.
|
| 7 |
- MiniMax-H3-Image-pruned_rank8_bf16.safetensors: the same, with the timestep modulation (time embedder and AdaLN projections) compressed to a rank-8 time table.
|
|
|
|
|
|
|
|
|
| 5 |
The following files were modified by DeepBeepMeep for WanGP from the released MiniMax-H3-Image checkpoint:
|
| 6 |
- MiniMax-H3-Image_bf16.safetensors: the 12 transformer shards merged into one file; the per-step output heads of pdd_fused_heads.safetensors stored in it as pdd_heads.video_out.*, replacing the unused final_layer.video_out. All other tensors are unchanged.
|
| 7 |
- MiniMax-H3-Image-pruned_rank8_bf16.safetensors: the same, with the timestep modulation (time embedder and AdaLN projections) compressed to a rank-8 time table.
|
| 8 |
+
- MiniMax-H3-Image_int8_convrot.safetensors and MiniMax-H3-Image-pruned_rank8_int8_convrot.safetensors: the two files above with the block matrices quantized to INT8 (ConvRot, Hadamard groups of 256, per-row FP32 scales) and the fused qkv rows grouped by projection.
|
| 9 |
+
- minimax_h3_image/minimax_h3_image_vae_fp16.safetensors: the released VAE in FP16, its post_quant_conv stored as a 1x1x1 3D convolution and the unused decoder mask token removed.
|
minimax_h3_image/minimax_h3_image_vae_fp16.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bbbf242ada140c05d353b0c6588d4a5641d74311f9c31139ed6f4f9517b0082a
|
| 3 |
+
size 1168503248
|