Alex995647 commited on
Commit
e8b58cc
·
verified ·
1 Parent(s): efd4f18

Add Spatial & Physics LoRA (clean, 3000 steps)

Browse files
.gitattributes CHANGED
@@ -34,3 +34,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  h3-realism-people/example_images/01_video.mp4 filter=lfs diff=lfs merge=lfs -text
 
 
 
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  h3-realism-people/example_images/01_video.mp4 filter=lfs diff=lfs merge=lfs -text
37
+ spatial-physics-lora-clean-3000-steps/example_images/01_video.mp4 filter=lfs diff=lfs merge=lfs -text
38
+ spatial-physics-lora-clean-3000-steps/example_images/02_video.mp4 filter=lfs diff=lfs merge=lfs -text
spatial-physics-lora-clean-3000-steps/example_images/01_poster.jpg ADDED
spatial-physics-lora-clean-3000-steps/example_images/01_video.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:111b798bfb050a69fd2546a5dcf5a683686937f6b6b293df867c8998e7fb8c77
3
+ size 4416854
spatial-physics-lora-clean-3000-steps/example_images/02_poster.jpg ADDED
spatial-physics-lora-clean-3000-steps/example_images/02_video.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:570dd176f6f9a7d41f3ca8951849b78f37ae3922834bcdbf1a8b78ff8cdcb00e
3
+ size 4663732
spatial-physics-lora-clean-3000-steps/info.txt ADDED
@@ -0,0 +1,145 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ═══════════════════════════════════════════════════════════════════════════════
2
+ SPATIAL & PHYSICS LORA (CLEAN, 3000 STEPS)
3
+ ═══════════════════════════════════════════════════════════════════════════════
4
+
5
+ ───────────────────────────────────────────────────────────────────────────────
6
+ IDENTITY
7
+ ───────────────────────────────────────────────────────────────────────────────
8
+ Base model MiniMax H3
9
+ Generation mode Text-to-Video (T2VA)
10
+ Checkpoint fl2va (drives T2VA / I2VA / L2VA / FL2VA)
11
+ Category Motion quality
12
+ Author Jojocodex
13
+ Source https://huggingface.co/Jojocodex/minimax-h3-spatial-physics-lora
14
+ Provenance HuggingFace · Jojocodex/minimax-h3-spatial-physics-lora · wushu_spatial_physics_clean_3000_pruned.safetensors
15
+
16
+ ───────────────────────────────────────────────────────────────────────────────
17
+ WEIGHTS
18
+ ───────────────────────────────────────────────────────────────────────────────
19
+ File wushu_spatial_physics_clean_3000_pruned.safetensors
20
+ Version main
21
+ Size 147.9 MB
22
+ Format SafeTensor / unknown precision
23
+ SHA256 7d14f3701560068e7004159c8b2a7278bd2dbfc9e5e3b60d0bc9aef6c049919d
24
+ Published 2026-08-19
25
+
26
+ ───────────────────────────────────────────────────────────────────────────────
27
+ TRIGGER WORDS
28
+ ───────────────────────────────────────────────────────────────────────────────
29
+ No trigger word. The LoRA applies as soon as it is loaded —
30
+ strength is the only way to dial the effect up or down.
31
+
32
+ ───────────────────────────────────────────────────────────────────────────────
33
+ RECOMMENDED SETTINGS
34
+ ───────────────────────────────────────────────────────────────────────────────
35
+ Taken verbatim from the author's description:
36
+
37
+ Frame count 90
38
+
39
+ ───────────────────────────────────────────────────────────────────────────────
40
+ ADOPTION & POPULARITY
41
+ ───────────────────────────────────────────────────────────────────────────────
42
+ Downloads 11,457
43
+ Rating 100% positive (100/100)
44
+ Comments 0
45
+ Published 16 days ago
46
+ Download rate 713.5/day since release
47
+
48
+ Adoption rank #6 of 114 in this collection (Top tier)
49
+ Momentum rank #14 of 114 by download rate (Trending · fresh release)
50
+
51
+ Adoption tier is the percentile of total downloads within this collection;
52
+ momentum is the percentile of downloads-per-day since release. Momentum is
53
+ ranked rather than measured against a fixed rate, because this base model
54
+ is itself new and every LoRA looks fast on an absolute scale. Both numbers
55
+ are CivitAI's, read on the scrape date at the foot of this file.
56
+
57
+ ───────────────────────────────────────────────────────────────────────────────
58
+ CHAINING / STACKING
59
+ ───────────────────────────────────────────────────────────────────────────────
60
+ Stacks with others not stated
61
+
62
+ The author gave no chaining guidance. The rules below are the
63
+ general MiniMax H3 ones, not anything specific to this LoRA.
64
+
65
+ General MiniMax H3 rules that apply here (see REFERENCE.txt):
66
+
67
+ • Load the LoRA with LoraLoaderModelOnly between UNETLoader and your
68
+ sampler; stacking means chaining several of those nodes in series.
69
+ • Stacks normally on top of a single acceleration/turbo LoRA; keep only
70
+ one distill LoRA active in the chain.
71
+ • Chain this only onto an fl2va checkpoint (T2VA / I2VA / L2VA / FL2VA).
72
+ Loading it against ref2va can degrade output silently rather than
73
+ erroring.
74
+
75
+ ───────────────────────────────────────────────────────────────────────────────
76
+ PROS
77
+ ───────────────────────────────────────────────────────────────────────────────
78
+ + Widely adopted — 11,457 downloads. [source: CivitAI stats]
79
+ + Positively received — 100 thumbs-up with no down-votes. [source:
80
+ CivitAI stats]
81
+ + Compact at 148 MB — cheap to stack with other LoRAs. [source: file
82
+ size]
83
+ + Author published concrete recommended settings. [source: description]
84
+
85
+ ───────────────────────────────────────────────────────────────────────────────
86
+ CONS & CAVEATS
87
+ ───────────────────────────────────────────────────────────────────────────────
88
+ - No trigger word — the LoRA is always on once loaded, so strength is the
89
+ only control. [source: model metadata]
90
+
91
+ ───────────────────────────────────────────────────────────────────────────────
92
+ COMPATIBILITY WARNING
93
+ ───────────────────────────────────────────────────────────────────────────────
94
+ MiniMax H3 ships two transformers that share tensor key names: the fl2va
95
+ checkpoint (driving T2VA, I2VA, L2VA and FL2VA) and the separate ref2va
96
+ checkpoint (driving Ref2VA). Loading a LoRA against the wrong family can
97
+ silently degrade output instead of raising an error, so confirm the family
98
+ above matches the checkpoint you have loaded. Likewise, LoRAs built for
99
+ full checkpoints often throw adaln_proj shape errors on pruned ones — match
100
+ the build to your checkpoint.
101
+
102
+ ───────────────────────────────────────────────────────────────────────────────
103
+ AUTHOR'S DESCRIPTION
104
+ ───────────────────────────────────────────────────────────────────────────────
105
+ Current retrained version of the spatial-reasoning and object-physics LoRA
106
+ for MiniMax-H3, teaching pure-object spatial relationships and physical
107
+ motion: collision, stacking, falling, occlusion, bouncing and momentum
108
+ transfer. The author states it is still in training/testing, so it adds a
109
+ modest physics boost on top of the base model rather than transforming it,
110
+ and it also helps fight scenes where bodies collide. Trained with ai-
111
+ toolkit at rank 16, 512 resolution, 90 frames @ 24fps; compatible with and
112
+ stackable on the MiniMax-H3 Turbo acceleration LoRA. Header confirms 208
113
+ rank-16 lora_A/lora_B pairs, converted by convert_h3_lora.py from
114
+ wushu_spatial_physics_clean_000003000 with 100 AdaLN adapter keys removed
115
+ for the pruned base.
116
+
117
+ ───────────────────────────────────────────────────────────────────────────────
118
+ EXAMPLE MEDIA
119
+ ───────────────────────────────────────────────────────────────────────────────
120
+ 2 preview(s) in ./example_images/ — community results for
121
+ this LoRA. Videos keep a matching _poster.jpg still frame.
122
+
123
+ No prompts are recorded below: CivitAI's public API flags these posts as
124
+ having generation metadata but does not return it, so the prompts cannot be
125
+ scraped. Open the model page to read them.
126
+
127
+ 01_video.mp4 (video, )
128
+ still: 01_poster.jpg
129
+ 02_video.mp4 (video, )
130
+ still: 02_poster.jpg
131
+
132
+ ───────────────────────────────────────────────────────────────────────────────
133
+ HOW TO USE (COMFYUI)
134
+ ───────────────────────────────────────────────────────────────────────────────
135
+ 1. Copy the .safetensors file into ComfyUI/models/loras/
136
+ 2. Load the matching H3 checkpoint with UNETLoader.
137
+ 3. Insert LoraLoaderModelOnly between the loader and your sampler.
138
+ 4. Set strength as listed above, then add any trigger word to the prompt.
139
+ 5. Keep CFG at 1.0 with BasicGuider unless you know you want real CFG.
140
+
141
+ ───────────────────────────────────────────────────────────────────────────────
142
+ Scraped by lorakit on 2026-09-04 23:06 UTC.
143
+ Stats and description are the author's; pros/cons are derived from the
144
+ evidence tagged beside each line.
145
+ ───────────────────────────────────────────────────────────────────────────────
spatial-physics-lora-clean-3000-steps/metadata.json ADDED
@@ -0,0 +1,112 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "name": "Spatial & Physics LoRA (clean, 3000 steps)",
3
+ "slug": "spatial-physics-lora-clean-3000-steps",
4
+ "base_model": "MiniMax H3",
5
+ "media_type": "video",
6
+ "model_id": 1288658020,
7
+ "version_id": 1288658021,
8
+ "version_name": "main",
9
+ "source": "huggingface",
10
+ "url": "https://huggingface.co/Jojocodex/minimax-h3-spatial-physics-lora",
11
+ "repo_id": "Jojocodex/minimax-h3-spatial-physics-lora",
12
+ "file_path": "wushu_spatial_physics_clean_3000_pruned.safetensors",
13
+ "author": "Jojocodex",
14
+ "category": {
15
+ "key": "motion",
16
+ "label": "Motion quality"
17
+ },
18
+ "modalities": [
19
+ "t2v"
20
+ ],
21
+ "primary_modality": "t2v",
22
+ "checkpoint_family": "fl2va",
23
+ "trigger_words": [],
24
+ "recommended_settings": {
25
+ "Frame count": "90"
26
+ },
27
+ "chaining": {
28
+ "author": {
29
+ "stacks": "not stated",
30
+ "notes": "",
31
+ "avoid": [],
32
+ "requires": [],
33
+ "order": "",
34
+ "strength": "",
35
+ "stated": false
36
+ },
37
+ "general_rules": [
38
+ "Load the LoRA with LoraLoaderModelOnly between UNETLoader and your sampler; stacking means chaining several of those nodes in series.",
39
+ "Stacks normally on top of a single acceleration/turbo LoRA; keep only one distill LoRA active in the chain.",
40
+ "Chain this only onto an fl2va checkpoint (T2VA / I2VA / L2VA / FL2VA). Loading it against ref2va can degrade output silently rather than erroring."
41
+ ]
42
+ },
43
+ "pros": [
44
+ "Widely adopted — 11,457 downloads. [source: CivitAI stats]",
45
+ "Positively received — 100 thumbs-up with no down-votes. [source: CivitAI stats]",
46
+ "Compact at 148 MB — cheap to stack with other LoRAs. [source: file size]",
47
+ "Author published concrete recommended settings. [source: description]"
48
+ ],
49
+ "cons": [
50
+ "No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata]"
51
+ ],
52
+ "file": {
53
+ "name": "wushu_spatial_physics_clean_3000_pruned.safetensors",
54
+ "size_bytes": 155109672,
55
+ "size_mb": 147.92,
56
+ "sha256": "7d14f3701560068e7004159c8b2a7278bd2dbfc9e5e3b60d0bc9aef6c049919d"
57
+ },
58
+ "stats": {
59
+ "downloads": 11457,
60
+ "thumbs_up": 100,
61
+ "thumbs_down": 0
62
+ },
63
+ "popularity": {
64
+ "downloads": 11457,
65
+ "thumbs_up": 100,
66
+ "thumbs_down": 0,
67
+ "comments": 0,
68
+ "approval": "100% positive (100/100)",
69
+ "age_days": 16.1,
70
+ "downloads_per_day": 713.48,
71
+ "rank_by_downloads": 6,
72
+ "rank_by_momentum": 14,
73
+ "of_total": 114,
74
+ "adoption_tier": "Top tier",
75
+ "momentum": "Trending · fresh release"
76
+ },
77
+ "selection_score": 12657.0,
78
+ "selection_reasons": [
79
+ "11,457 downloads",
80
+ "100 thumbs-up",
81
+ "top-ranked for its category"
82
+ ],
83
+ "example_media": [
84
+ {
85
+ "file": "01_video.mp4",
86
+ "type": "video",
87
+ "resolution": "",
88
+ "prompt": "",
89
+ "source_path": "video/with lora.mp4",
90
+ "poster": "01_poster.jpg"
91
+ },
92
+ {
93
+ "file": "02_video.mp4",
94
+ "type": "video",
95
+ "resolution": "",
96
+ "prompt": "",
97
+ "source_path": "video/without lora.mp4",
98
+ "poster": "02_poster.jpg"
99
+ }
100
+ ],
101
+ "tags": [
102
+ "minimax-h3",
103
+ "lora",
104
+ "text-to-video",
105
+ "physics",
106
+ "spatial-reasoning",
107
+ "object-motion",
108
+ "video-generation"
109
+ ],
110
+ "scraped_at": "2026-09-04T23:06:25+00:00",
111
+ "path": "minimax-h3/spatial-physics-lora-clean-3000-steps"
112
+ }
spatial-physics-lora-clean-3000-steps/wushu_spatial_physics_clean_3000_pruned.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7d14f3701560068e7004159c8b2a7278bd2dbfc9e5e3b60d0bc9aef6c049919d
3
+ size 155109672