Hugging Face's logo

Kijai
/
MiniMax-H3-experimental

Model card Files
xet
 Community
60

MiniMax-H3-exper

#38
by willowoods - opened
base: refs/heads/main
←
from: refs/pr/38
Discussion Files changed
+9
-84
Files changed (18) hide show
  1. README.md +9 -39
  2. {model_patches → controlnet}/minimax_h3_fun_controlnet_union_pruned_bf16.safetensors +0 -0
  3. {model_patches → controlnet}/minimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors +0 -0
  4. loras/minimax_h3_4step_lora_flashgen_v1.0_768p_fl2va_pruned_avg_rank_13_bf16.safetensors +0 -3
  5. loras/minimax_h3_DMAD_4step_full_lora_avg_rank_39_bf16.safetensors +0 -3
  6. loras/minimax_h3_ELM_longlive_plug_4step_lora_bf16.safetensors +0 -3
  7. loras/minimax_h3_open_regenerate_2k_refiner_rank_16_bf16.safetensors +0 -3
  8. loras/minimax_h3_open_regenerate_2k_refiner_rank_384_bf16.safetensors +0 -3
  9. loras/minimax_h3_pdmd_2step_lora_avg_rank_38_bf16.safetensors +0 -3
  10. loras/minimax_h3_pdmd_4step_lora_avg_rank_57_bf16.safetensors +0 -3
  11. loras/minimax_h3_singularity_v2_lora_avg_rank_351_bf16.safetensors +0 -3
  12. loras/minimax_h3_taomate_fl2va_3step_avg_rank_19_bf16.safetensors +0 -3
  13. minimax_h3_lightvae_lite_encoder_fp16.safetensors +0 -3
  14. minimax_h3_lightvae_lite_vae_int8_convrot.safetensors +0 -3
  15. minimax_h3_lynnreal_light_vae_int8_convrot.safetensors +0 -3
  16. model_patches/minimax_h3_fun_controlnet_union_2.0_pruned_bf16.safetensors +0 -3
  17. model_patches/minimax_h3_fun_controlnet_union_2.0_pruned_int8_convrot.safetensors +0 -3
  18. wip/minimax_h3_ref2va_pruned_w6a8_g32.safetensors +0 -3
README.md CHANGED
@@ -1,12 +1,14 @@
1
- w4a8 is a format using 4-bit weights and int8-convrot activations
 
 
2
 
3
  Requires ComfyUI 0.31.0
4
 
5
  ---
6
 
7
- int8_convrot VAE needs ComfyUI 0.36.0 for maximum speed benefit
8
 
9
- It speeds up VAE decode times by ~2x
10
 
11
  ---
12
 
@@ -14,42 +16,10 @@ ref lora is the difference between fl2va and ref2va, completely experimental, I
14
 
15
  ---
16
 
17
- w6a8 is a format using 6-bit weights and int8-convrot activation, work in progress
18
-
19
- ## minimax_h3_ref2va_pruned_w6a8_g32.safetensor
20
-
21
- https://github.com/Comfy-Org/comfy-kitchen/pull/191
22
-
23
- ---
24
-
25
- ## minimax_h3_lynnreal_light_vae_int8_convrot.safetensors
26
-
27
- https://huggingface.co/stdstu123/LynnReal-Onmi-light-vae
28
-
29
- int8-convrot quant of the pruned light VAE, this is about 1.3x faster for decoding.
30
-
31
- ---
32
-
33
- ## PDMD loras
34
-
35
- Converted and rank reduced from:
36
-
37
- https://huggingface.co/pdmd2026/pdmd_2NFE_lora
38
-
39
- https://huggingface.co/pdmd2026/pdmd_4NFE_lora
40
-
41
- ## DMAD lora
42
-
43
- https://huggingface.co/ZhengmingYu/DMAD
44
-
45
- ## Singularity lora extraction
46
-
47
- `minimax_h3_singularity_v2_lora_avg_rank_351_bf16.safetensors`
48
-
49
- Extracted from https://huggingface.co/WarmBloodAban/Minimax-h3_Singularity/blob/main/Minimax-h3_Singularity_ref2va_v2_int8.safetensors
50
 
51
- Note that even at this rank extraction of full finetune only captures maybe 40-50% of it.
52
 
53
- ## minimax_h3_open_regenerate_2k_refiner
54
 
55
- The refiner part ONLY, doesn't include the LMS or the lightx2v turbo.
 
1
+ w4a8 is experimental format using 4bit weights and int8-convrot activations, at testing phase, further info here:
2
+
3
+ https://github.com/Comfy-Org/comfy-kitchen/pull/90
4
 
5
  Requires ComfyUI 0.31.0
6
 
7
  ---
8
 
9
+ int8_convrot VAE needs ComfyUI 0.31.0 or you get black outputs
10
 
11
+ It speeds up VAE decode times by ~1.5x
12
 
13
  ---
14
 
 
16
 
17
  ---
18
 
19
+ minimax_h3_fastvideo_vsa_datafree_1300step_4step_int8_convrot.safetensors
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
20
 
21
+ Currently for testing with these PRs:
22
 
23
+ https://github.com/Comfy-Org/ComfyUI/pull/15958
24
 
25
+ https://github.com/Comfy-Org/comfy-kitchen/pull/117
{model_patches → controlnet}/minimax_h3_fun_controlnet_union_pruned_bf16.safetensors RENAMED
File without changes
{model_patches → controlnet}/minimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors RENAMED
File without changes
loras/minimax_h3_4step_lora_flashgen_v1.0_768p_fl2va_pruned_avg_rank_13_bf16.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:1254d62240b1f2f0f2376676338dce40d0464d5f1c6069fc0541c1e8f8f7b067
3
- size 185865134
 
 
 
 
loras/minimax_h3_DMAD_4step_full_lora_avg_rank_39_bf16.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:ddfdd94f4c9a0f32bf3649bb8e98ca20db572e4db6bcd9cca546a5b33ea32d11
3
- size 366647000
 
 
 
 
loras/minimax_h3_ELM_longlive_plug_4step_lora_bf16.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:4a1058d29eec07a2088c51ac38863fdbcab744127aa313b2bb4845c51b4df214
3
- size 1956171824
 
 
 
 
loras/minimax_h3_open_regenerate_2k_refiner_rank_16_bf16.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:add559cd92e2b17b42ac080516b476df120903e485fc98add4538e0217820fbd
3
- size 155133472
 
 
 
 
loras/minimax_h3_open_regenerate_2k_refiner_rank_384_bf16.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:4bf30291343eef5e51ee867ba37dd31bce3133d1da1ace1a245a7130b88818c2
3
- size 5868409512
 
 
 
 
loras/minimax_h3_pdmd_2step_lora_avg_rank_38_bf16.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:a17f3eb45ab16be56fce79a522604a8ec1ce41d732022ad398c5908b230eec55
3
- size 381015376
 
 
 
 
loras/minimax_h3_pdmd_4step_lora_avg_rank_57_bf16.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:b5c62eb228a4f604890566ad698713703ed221302e0df1c067725c39f11efc37
3
- size 558735392
 
 
 
 
loras/minimax_h3_singularity_v2_lora_avg_rank_351_bf16.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:dcba5922efc0b211c5bba9ea5a530917ac59b37c46b856d2e42a35428361f3c8
3
- size 3300296560
 
 
 
 
loras/minimax_h3_taomate_fl2va_3step_avg_rank_19_bf16.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:e6f86322a0158dcf804b106e784a9ec1a078b1522c161143627364edc08a8bd9
3
- size 176723664
 
 
 
 
minimax_h3_lightvae_lite_encoder_fp16.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:171b1111415878ca34e4fd122fdda5638cd31e522ec8693edbb7aee45290b7fd
3
- size 203172976
 
 
 
 
minimax_h3_lightvae_lite_vae_int8_convrot.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:d9bde765b71ac35ee6531a32546c7ad2d50b89e5a812e1b0962b20add1ee6867
3
- size 1036943392
 
 
 
 
minimax_h3_lynnreal_light_vae_int8_convrot.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:f11f8b9b96c9dd22b36bc6c18464b6929c207d247e6d865d90fe31588e58e7e2
3
- size 2137493384
 
 
 
 
model_patches/minimax_h3_fun_controlnet_union_2.0_pruned_bf16.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:cd48bdc11b1f1a3b934746c898d25123754b335c13d37d5c421f8373079282b4
3
- size 8382288792
 
 
 
 
model_patches/minimax_h3_fun_controlnet_union_2.0_pruned_int8_convrot.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:890af58bb350f0c2f6c409b1c67ea8f4196654297d037769b72de1e6f504c612
3
- size 4531220608
 
 
 
 
wip/minimax_h3_ref2va_pruned_w6a8_g32.safetensors DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:565519608d06ceb206859522a6514c4470951ae59a761f23f557b5f81900db01
3
- size 16753960844