Hugging Face's logo

aipracticecafe
/
dual-stream-dit

Text-to-Image
Diffusers
PyTorch
flow-matching
rectified-flow
dit
dual-stream-dit
sprint
anime
Model card Files
xet
 Community
aipracticecafe commited on
Commit
3363c19
·
verified ·
1 Parent(s): 0d24fb9

Upload folder using huggingface_hub

Browse files
Files changed (10) hide show
  1. epoch_204_step_281281/config.json +110 -0
  2. epoch_204_step_281281/optimizer.pt +3 -0
  3. epoch_204_step_281281/scheduler.pt +3 -0
  4. epoch_204_step_281281/training_state.pt +3 -0
  5. epoch_204_step_281281/unet.safetensors +3 -0
  6. epoch_206_step_286000/config.json +110 -0
  7. epoch_206_step_286000/optimizer.pt +3 -0
  8. epoch_206_step_286000/scheduler.pt +3 -0
  9. epoch_206_step_286000/training_state.pt +3 -0
  10. epoch_206_step_286000/unet.safetensors +3 -0
epoch_204_step_281281/config.json ADDED
@@ -0,0 +1,110 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_class_name": "sprint_dual",
3
+ "model_type": "sprint_dual",
4
+ "in_channels": 32,
5
+ "hidden_size": 768,
6
+ "depth": 13,
7
+ "num_heads": 12,
8
+ "encoder_depth": 2,
9
+ "decoder_depth": 2,
10
+ "drop_ratio": 0.0,
11
+ "drop_target": "image",
12
+ "residual_type": "concat_linear",
13
+ "hf_text_encoder": "LiquidAI/LFM2.5-230M",
14
+ "hf_vae": "kaiyuyue/FLUX.2-dev-vae",
15
+ "vae_mean": 0.05106564902828182,
16
+ "vae_std": 1.69915383699,
17
+ "resume_from_checkpoint": "results/discrete_rope/epoch_194_step_265389",
18
+ "cfg_mask_prob": 0.0,
19
+ "skip_checkpointing_layers": 0,
20
+ "use_random_drop": false,
21
+ "vae_batch_size": 8,
22
+ "use_calibrated_spatial": true,
23
+ "train": {
24
+ "batch_size": 32,
25
+ "lr": 0.00025,
26
+ "epochs": 205,
27
+ "gradient_accumulation_steps": 8,
28
+ "dtype": "fp32",
29
+ "use_checkpointing": true,
30
+ "train_te": false,
31
+ "wd": 0.01,
32
+ "warmup": 0.04,
33
+ "min_lr_ratio": 0.01,
34
+ "reset_scheduler": true,
35
+ "reset_optimizer": false,
36
+ "use_cos_scheduler": false,
37
+ "objective": "flow_matching",
38
+ "timestep_fn": "logit-normal",
39
+ "shift": 2.0,
40
+ "snr_gamma": 5.0,
41
+ "input_perturb": 0.1,
42
+ "cfg_dropout_prob": 0.1,
43
+ "seed": 232,
44
+ "use_bitsandbytes": true,
45
+ "train_only_output": false,
46
+ "skip_save_n_times": 4,
47
+ "run_stability_check": false,
48
+ "world_size": 1,
49
+ "use_ot": false,
50
+ "use_ema": false,
51
+ "ema_decay": 0.999,
52
+ "use_kahan_sum": false,
53
+ "use_cpu_accumulator": false,
54
+ "compile_model": true,
55
+ "gpu_peak_tflops": 165.2
56
+ },
57
+ "data": {
58
+ "dataset_type": "streaming",
59
+ "h5_path": "data",
60
+ "streaming_dataset_name": "aipracticecafe/curated-danbooru-2026-512px-flux2-vae",
61
+ "num_workers": 4,
62
+ "prefetch_factor": 4,
63
+ "load_into_ram": true,
64
+ "tag_dropout": 0.1,
65
+ "use_shift_crop": true,
66
+ "resolution": 512,
67
+ "buffer_size": 2500,
68
+ "cache_latents_to_ram": false,
69
+ "is_latent": true,
70
+ "precompute_num_workers": 2,
71
+ "precompute_prefetch_factor": 2
72
+ },
73
+ "models": {
74
+ "model_type": "sprint_dual",
75
+ "resume_from_checkpoint": "results/discrete_rope/epoch_194_step_265389",
76
+ "output_head_path": null,
77
+ "hf_text_encoder": "LiquidAI/LFM2.5-230M",
78
+ "hidden_size": 768,
79
+ "depth": 13,
80
+ "num_heads": 12,
81
+ "encoder_depth": 2,
82
+ "decoder_depth": 2,
83
+ "drop_ratio": 0.0,
84
+ "drop_target": "image",
85
+ "residual_type": "concat_linear",
86
+ "cfg_mask_prob": 0.0,
87
+ "skip_checkpointing_layers": 0,
88
+ "use_random_drop": false,
89
+ "hf_vae": "kaiyuyue/FLUX.2-dev-vae",
90
+ "vae_mean": 0.05106564902828182,
91
+ "vae_std": 1.69915383699,
92
+ "vae_batch_size": 8,
93
+ "in_channels": 32,
94
+ "use_calibrated_spatial": true
95
+ },
96
+ "paths": {
97
+ "models": "models"
98
+ },
99
+ "logging": {
100
+ "project": "dual-dit",
101
+ "wandb_entity": "aipracticecafe-aipracticecafe",
102
+ "save_interval": 6500,
103
+ "hf_repo": "dual-stream-dit-spatial-rope",
104
+ "wandb_run_name_prefix": "4090-sprint-512px-latents-rope-swap-fine"
105
+ },
106
+ "sampling": {
107
+ "interval": 1300,
108
+ "config_file": "sample.toml"
109
+ }
110
+ }
epoch_204_step_281281/optimizer.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4f0c4391f885610bd3e23c6d47fe9c631056aed42bff7c9db1b0a10f7140ec83
3
+ size 419905931
epoch_204_step_281281/scheduler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85e2d6979d4a255e96df191b71c444fff329d8db1ad206e8ee38f49edcf914e1
3
+ size 1721
epoch_204_step_281281/training_state.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:927b861d2ef6485870637f5586571389cd06b9e02d8bb4fbb4ff30e0260ca747
3
+ size 1303
epoch_204_step_281281/unet.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c53d6620989681b02994146ed8db09823608cfe75d9ff224777b69fb4a9f5ee3
3
+ size 425266512
epoch_206_step_286000/config.json ADDED
@@ -0,0 +1,110 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_class_name": "sprint_dual",
3
+ "model_type": "sprint_dual",
4
+ "in_channels": 32,
5
+ "hidden_size": 768,
6
+ "depth": 13,
7
+ "num_heads": 12,
8
+ "encoder_depth": 2,
9
+ "decoder_depth": 2,
10
+ "drop_ratio": 0.0,
11
+ "drop_target": "image",
12
+ "residual_type": "concat_linear",
13
+ "hf_text_encoder": "LiquidAI/LFM2.5-230M",
14
+ "hf_vae": "kaiyuyue/FLUX.2-dev-vae",
15
+ "vae_mean": 0.05106564902828182,
16
+ "vae_std": 1.69915383699,
17
+ "resume_from_checkpoint": "results/discrete_rope/epoch_204_step_281281",
18
+ "cfg_mask_prob": 0.0,
19
+ "skip_checkpointing_layers": 4,
20
+ "use_random_drop": false,
21
+ "vae_batch_size": 8,
22
+ "use_calibrated_spatial": true,
23
+ "train": {
24
+ "batch_size": 32,
25
+ "lr": 0.00025,
26
+ "epochs": 210,
27
+ "gradient_accumulation_steps": 8,
28
+ "dtype": "fp32",
29
+ "use_checkpointing": true,
30
+ "train_te": false,
31
+ "wd": 0.01,
32
+ "warmup": 0.04,
33
+ "min_lr_ratio": 0.01,
34
+ "reset_scheduler": true,
35
+ "reset_optimizer": false,
36
+ "use_cos_scheduler": true,
37
+ "objective": "flow_matching",
38
+ "timestep_fn": "logit-normal",
39
+ "shift": 2.0,
40
+ "snr_gamma": 5.0,
41
+ "input_perturb": 0.1,
42
+ "cfg_dropout_prob": 0.1,
43
+ "seed": 232,
44
+ "use_bitsandbytes": true,
45
+ "train_only_output": false,
46
+ "skip_save_n_times": 4,
47
+ "run_stability_check": false,
48
+ "world_size": 1,
49
+ "use_ot": false,
50
+ "use_ema": false,
51
+ "ema_decay": 0.999,
52
+ "use_kahan_sum": false,
53
+ "use_cpu_accumulator": false,
54
+ "compile_model": true,
55
+ "gpu_peak_tflops": 165.2
56
+ },
57
+ "data": {
58
+ "dataset_type": "streaming",
59
+ "h5_path": "data",
60
+ "streaming_dataset_name": "aipracticecafe/curated-danbooru-2026-512px-flux2-vae",
61
+ "num_workers": 4,
62
+ "prefetch_factor": 4,
63
+ "load_into_ram": true,
64
+ "tag_dropout": 0.1,
65
+ "use_shift_crop": true,
66
+ "resolution": 512,
67
+ "buffer_size": 2500,
68
+ "cache_latents_to_ram": false,
69
+ "is_latent": true,
70
+ "precompute_num_workers": 2,
71
+ "precompute_prefetch_factor": 2
72
+ },
73
+ "models": {
74
+ "model_type": "sprint_dual",
75
+ "resume_from_checkpoint": "results/discrete_rope/epoch_204_step_281281",
76
+ "output_head_path": null,
77
+ "hf_text_encoder": "LiquidAI/LFM2.5-230M",
78
+ "hidden_size": 768,
79
+ "depth": 13,
80
+ "num_heads": 12,
81
+ "encoder_depth": 2,
82
+ "decoder_depth": 2,
83
+ "drop_ratio": 0.0,
84
+ "drop_target": "image",
85
+ "residual_type": "concat_linear",
86
+ "cfg_mask_prob": 0.0,
87
+ "skip_checkpointing_layers": 4,
88
+ "use_random_drop": false,
89
+ "hf_vae": "kaiyuyue/FLUX.2-dev-vae",
90
+ "vae_mean": 0.05106564902828182,
91
+ "vae_std": 1.69915383699,
92
+ "vae_batch_size": 8,
93
+ "in_channels": 32,
94
+ "use_calibrated_spatial": true
95
+ },
96
+ "paths": {
97
+ "models": "models"
98
+ },
99
+ "logging": {
100
+ "project": "dual-dit",
101
+ "wandb_entity": "aipracticecafe-aipracticecafe",
102
+ "save_interval": 13000,
103
+ "hf_repo": "dual-stream-dit-spatial-rope",
104
+ "wandb_run_name_prefix": "4090-sprint-512px-latents-rope-swap-decay"
105
+ },
106
+ "sampling": {
107
+ "interval": 1300,
108
+ "config_file": "sample.toml"
109
+ }
110
+ }
epoch_206_step_286000/optimizer.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6974d541934a43114124f173b40c071063fe731ef467c46d568644b7640116ab
3
+ size 419905931
epoch_206_step_286000/scheduler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2e3fc698f84a0add37eec68cac292d02d1dd8f33e1cc7e7e4d67e53b8bde5159
3
+ size 1721
epoch_206_step_286000/training_state.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bc3595568800723380f107da35a25669c46cf946e1a994071960a612add83132
3
+ size 1303
epoch_206_step_286000/unet.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a3e5a9dc1dab2619d9e540ba8c7e8c0e92c55c8077fbbb3ad8dcbcc093394323
3
+ size 425266512