Add files using upload-large-folder tool
Browse files- geneval_sd3_dreamsim_fd_sim_log/step_10/model.safetensors +3 -0
- geneval_sd3_dreamsim_fd_sim_log/step_20/model.safetensors +3 -0
- geneval_sd3_dreamsim_fd_sim_log/step_30/model.safetensors +3 -0
- geneval_sd3_fd3_kl_base_A1_ema05/config.yaml +96 -0
- geneval_sd3_fd3_kl_base_F_fdcap/config.yaml +96 -0
- geneval_sd3_fd3_kl_dreamsim14_phi015_rm150/step_10/model.safetensors +3 -0
- geneval_sd3_fd3_kl_dreamsim14_phi015_rm150/step_120/model.safetensors +3 -0
- geneval_sd3_fd3_kl_dreamsim14_phi015_rm150/step_70/model.safetensors +3 -0
- geneval_sd3_fd3_kl_dreamsim14_phi015_rm150/step_80/model.safetensors +3 -0
- geneval_sd3_fd3_kl_log/config.yaml +94 -0
- geneval_sd3_fd3_kl_phi015/config.yaml +94 -0
- geneval_sd3_fd3_kl_phi030/config.yaml +94 -0
- geneval_sd3_fd3_phi015/step_10/model.safetensors +3 -0
- geneval_sd3_fd3_phi015/step_20/model.safetensors +3 -0
- geneval_sd3_fd3_phi015/step_30/model.safetensors +3 -0
- geneval_sd3_fd3_phi015/step_40/model.safetensors +3 -0
geneval_sd3_dreamsim_fd_sim_log/step_10/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8071e8c38c61e663e886545dce656d421648cca84e3e9d024c158eaec32b5fda
|
| 3 |
+
size 5617919800
|
geneval_sd3_dreamsim_fd_sim_log/step_20/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:46d9d09cb68be70398431d3cf78bcb4cb441112278cb7399e4198d02402e6ce7
|
| 3 |
+
size 5617919800
|
geneval_sd3_dreamsim_fd_sim_log/step_30/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:25f26889be53b9ee8f4fc7531023e61e6a32ef5fbcc59d5834f1f3a3b8e29cbb
|
| 3 |
+
size 5617919800
|
geneval_sd3_fd3_kl_base_A1_ema05/config.yaml
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
beta1: 0.9
|
| 2 |
+
beta2: 0.99
|
| 3 |
+
dreamsim_batch_size: 12
|
| 4 |
+
dreamsim_cache_dir: models/dreamsim
|
| 5 |
+
dreamsim_eval_guidance_scale: 2.0
|
| 6 |
+
dreamsim_eval_num_prompts: 32
|
| 7 |
+
dreamsim_eval_samples_per_prompt: 8
|
| 8 |
+
dreamsim_gate: std
|
| 9 |
+
dreamsim_log_every_n_epochs: 1
|
| 10 |
+
dreamsim_log_types:
|
| 11 |
+
- dino_vitb16_patch
|
| 12 |
+
- clip_vitb32
|
| 13 |
+
- open_clip_vitb32
|
| 14 |
+
- ensemble
|
| 15 |
+
dreamsim_scale: 0.0
|
| 16 |
+
dreamsim_train_type: dino_vitb16_patch
|
| 17 |
+
ema_decay: 0.5
|
| 18 |
+
ema_warmup_rate: 0.01
|
| 19 |
+
eval_every_n_epochs: 10
|
| 20 |
+
eval_guidance_scales:
|
| 21 |
+
- 1.0
|
| 22 |
+
- 2.0
|
| 23 |
+
- 3.0
|
| 24 |
+
- 4.5
|
| 25 |
+
eval_prompt_fraction: 1.0
|
| 26 |
+
fd_allow_small_reference: false
|
| 27 |
+
fd_batch_size: 12
|
| 28 |
+
fd_cache_dir: models/fd_stats
|
| 29 |
+
fd_clamp: 10.0
|
| 30 |
+
fd_cov_source: residual
|
| 31 |
+
fd_cov_weight: 1.0
|
| 32 |
+
fd_eig_floor: 0.0
|
| 33 |
+
fd_ema_beta: 0.95
|
| 34 |
+
fd_encoder_cache_dir: models/fd_encoders
|
| 35 |
+
fd_init_epochs: 25
|
| 36 |
+
fd_log_every_n_epochs: 1
|
| 37 |
+
fd_log_types:
|
| 38 |
+
- dinov2_vitl14
|
| 39 |
+
- inception_v3
|
| 40 |
+
- mae_vitl16
|
| 41 |
+
fd_mean_weight: 1.0
|
| 42 |
+
fd_norm_eps: null
|
| 43 |
+
fd_normalize: null_ratio
|
| 44 |
+
fd_realized_max: null
|
| 45 |
+
fd_reference_seed: null
|
| 46 |
+
fd_sample_clip: null
|
| 47 |
+
fd_scale: 0.0
|
| 48 |
+
fd_shrinkage: 0.02
|
| 49 |
+
fd_train_types:
|
| 50 |
+
- convnextv2_base
|
| 51 |
+
- siglip2_so400m
|
| 52 |
+
- clip_vitl14
|
| 53 |
+
fd_train_weights:
|
| 54 |
+
- 1.0
|
| 55 |
+
- 1.0
|
| 56 |
+
- 1.0
|
| 57 |
+
grad_norm_monitor_every_n_epochs: 1
|
| 58 |
+
integral_kl_timestep_distribution: uniform
|
| 59 |
+
integral_kl_weight: 0.0
|
| 60 |
+
lora_alpha: 64
|
| 61 |
+
lora_rank: 32
|
| 62 |
+
loss_batch_size: 16
|
| 63 |
+
lr: 0.0003
|
| 64 |
+
min_timesteps: 400
|
| 65 |
+
mixed_precision: fp16
|
| 66 |
+
model_name: stabilityai/stable-diffusion-3.5-medium
|
| 67 |
+
num_eval_inference_steps: 40
|
| 68 |
+
num_integral_kl_timesteps: 0
|
| 69 |
+
num_loss_targets_per_sample: 8
|
| 70 |
+
num_optim_steps_per_epoch: 1
|
| 71 |
+
num_prompts_per_epoch: 48
|
| 72 |
+
num_samples_per_prompt: 24
|
| 73 |
+
num_train_inference_steps: 20
|
| 74 |
+
prompts: geneval
|
| 75 |
+
resolution: 512
|
| 76 |
+
reverse_kl_weight: 1.0
|
| 77 |
+
reward_batch_size: 12
|
| 78 |
+
reward_multiplier: 100
|
| 79 |
+
rewards:
|
| 80 |
+
- Geneval
|
| 81 |
+
sampling_batch_size: 32
|
| 82 |
+
scale_rewards: epoch
|
| 83 |
+
seed: 0
|
| 84 |
+
sft_loss_weight: 0.0
|
| 85 |
+
sft_loss_weight_fn: null
|
| 86 |
+
sft_num_loss_targets_per_sample: 8
|
| 87 |
+
sft_num_samples_per_prompt: 24
|
| 88 |
+
sft_ref_cache_dir: null
|
| 89 |
+
timestep_sampling: weighted
|
| 90 |
+
train_guidance_scale: 2.0
|
| 91 |
+
validation_rewards:
|
| 92 |
+
- ImageReward
|
| 93 |
+
wandb_entity: jaylee2000-kaist
|
| 94 |
+
wandb_project: ram
|
| 95 |
+
wandb_run_name: null
|
| 96 |
+
weight_decay: 0.0001
|
geneval_sd3_fd3_kl_base_F_fdcap/config.yaml
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
beta1: 0.9
|
| 2 |
+
beta2: 0.99
|
| 3 |
+
dreamsim_batch_size: 12
|
| 4 |
+
dreamsim_cache_dir: models/dreamsim
|
| 5 |
+
dreamsim_eval_guidance_scale: 2.0
|
| 6 |
+
dreamsim_eval_num_prompts: 32
|
| 7 |
+
dreamsim_eval_samples_per_prompt: 8
|
| 8 |
+
dreamsim_gate: std
|
| 9 |
+
dreamsim_log_every_n_epochs: 1
|
| 10 |
+
dreamsim_log_types:
|
| 11 |
+
- dino_vitb16_patch
|
| 12 |
+
- clip_vitb32
|
| 13 |
+
- open_clip_vitb32
|
| 14 |
+
- ensemble
|
| 15 |
+
dreamsim_scale: 0.0
|
| 16 |
+
dreamsim_train_type: dino_vitb16_patch
|
| 17 |
+
ema_decay: 0.9
|
| 18 |
+
ema_warmup_rate: 0.01
|
| 19 |
+
eval_every_n_epochs: 10
|
| 20 |
+
eval_guidance_scales:
|
| 21 |
+
- 1.0
|
| 22 |
+
- 2.0
|
| 23 |
+
- 3.0
|
| 24 |
+
- 4.5
|
| 25 |
+
eval_prompt_fraction: 1.0
|
| 26 |
+
fd_allow_small_reference: false
|
| 27 |
+
fd_batch_size: 12
|
| 28 |
+
fd_cache_dir: models/fd_stats
|
| 29 |
+
fd_clamp: 10.0
|
| 30 |
+
fd_cov_source: residual
|
| 31 |
+
fd_cov_weight: 1.0
|
| 32 |
+
fd_eig_floor: 0.0
|
| 33 |
+
fd_ema_beta: 0.95
|
| 34 |
+
fd_encoder_cache_dir: models/fd_encoders
|
| 35 |
+
fd_init_epochs: 25
|
| 36 |
+
fd_log_every_n_epochs: 1
|
| 37 |
+
fd_log_types:
|
| 38 |
+
- dinov2_vitl14
|
| 39 |
+
- inception_v3
|
| 40 |
+
- mae_vitl16
|
| 41 |
+
fd_mean_weight: 1.0
|
| 42 |
+
fd_norm_eps: null
|
| 43 |
+
fd_normalize: null_ratio
|
| 44 |
+
fd_realized_max: 0.1
|
| 45 |
+
fd_reference_seed: null
|
| 46 |
+
fd_sample_clip: 3.0
|
| 47 |
+
fd_scale: 0.06939
|
| 48 |
+
fd_shrinkage: 0.02
|
| 49 |
+
fd_train_types:
|
| 50 |
+
- convnextv2_base
|
| 51 |
+
- siglip2_so400m
|
| 52 |
+
- clip_vitl14
|
| 53 |
+
fd_train_weights:
|
| 54 |
+
- 1.412
|
| 55 |
+
- 0.662
|
| 56 |
+
- 1.278
|
| 57 |
+
grad_norm_monitor_every_n_epochs: 1
|
| 58 |
+
integral_kl_timestep_distribution: uniform
|
| 59 |
+
integral_kl_weight: 0.0
|
| 60 |
+
lora_alpha: 64
|
| 61 |
+
lora_rank: 32
|
| 62 |
+
loss_batch_size: 16
|
| 63 |
+
lr: 0.0003
|
| 64 |
+
min_timesteps: 400
|
| 65 |
+
mixed_precision: fp16
|
| 66 |
+
model_name: stabilityai/stable-diffusion-3.5-medium
|
| 67 |
+
num_eval_inference_steps: 40
|
| 68 |
+
num_integral_kl_timesteps: 0
|
| 69 |
+
num_loss_targets_per_sample: 8
|
| 70 |
+
num_optim_steps_per_epoch: 1
|
| 71 |
+
num_prompts_per_epoch: 48
|
| 72 |
+
num_samples_per_prompt: 24
|
| 73 |
+
num_train_inference_steps: 20
|
| 74 |
+
prompts: geneval
|
| 75 |
+
resolution: 512
|
| 76 |
+
reverse_kl_weight: 1.0
|
| 77 |
+
reward_batch_size: 12
|
| 78 |
+
reward_multiplier: 100
|
| 79 |
+
rewards:
|
| 80 |
+
- Geneval
|
| 81 |
+
sampling_batch_size: 32
|
| 82 |
+
scale_rewards: epoch
|
| 83 |
+
seed: 0
|
| 84 |
+
sft_loss_weight: 0.0
|
| 85 |
+
sft_loss_weight_fn: null
|
| 86 |
+
sft_num_loss_targets_per_sample: 8
|
| 87 |
+
sft_num_samples_per_prompt: 24
|
| 88 |
+
sft_ref_cache_dir: null
|
| 89 |
+
timestep_sampling: weighted
|
| 90 |
+
train_guidance_scale: 2.0
|
| 91 |
+
validation_rewards:
|
| 92 |
+
- ImageReward
|
| 93 |
+
wandb_entity: jaylee2000-kaist
|
| 94 |
+
wandb_project: ram
|
| 95 |
+
wandb_run_name: null
|
| 96 |
+
weight_decay: 0.0001
|
geneval_sd3_fd3_kl_dreamsim14_phi015_rm150/step_10/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:37caf505a4ef45da4e4cfdafdb0b12d0644bf6b731ea05414706221bbf106596
|
| 3 |
+
size 5617919800
|
geneval_sd3_fd3_kl_dreamsim14_phi015_rm150/step_120/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ab82bea12ae90ab6136b7bf2c3e9a4089c02667e90ffce6c66309eeb37c1bf78
|
| 3 |
+
size 5617919800
|
geneval_sd3_fd3_kl_dreamsim14_phi015_rm150/step_70/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f32f094cceabafec6ffeeda8d904a3f2e644f54a519a505ba0e7c1d0281e3be7
|
| 3 |
+
size 5617919800
|
geneval_sd3_fd3_kl_dreamsim14_phi015_rm150/step_80/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:35ee56b02cc44e3fe738fb11af1ea71f33840eb0ef23938e0953fafc8b0a0bb9
|
| 3 |
+
size 5617919800
|
geneval_sd3_fd3_kl_log/config.yaml
ADDED
|
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
beta1: 0.9
|
| 2 |
+
beta2: 0.99
|
| 3 |
+
dreamsim_batch_size: 12
|
| 4 |
+
dreamsim_cache_dir: models/dreamsim
|
| 5 |
+
dreamsim_eval_guidance_scale: 2.0
|
| 6 |
+
dreamsim_eval_num_prompts: 32
|
| 7 |
+
dreamsim_eval_samples_per_prompt: 8
|
| 8 |
+
dreamsim_gate: std
|
| 9 |
+
dreamsim_log_every_n_epochs: 1
|
| 10 |
+
dreamsim_log_types:
|
| 11 |
+
- dino_vitb16_patch
|
| 12 |
+
- clip_vitb32
|
| 13 |
+
- open_clip_vitb32
|
| 14 |
+
- ensemble
|
| 15 |
+
dreamsim_scale: 0.0
|
| 16 |
+
dreamsim_train_type: dino_vitb16_patch
|
| 17 |
+
ema_decay: 0.9
|
| 18 |
+
ema_warmup_rate: 0.01
|
| 19 |
+
eval_every_n_epochs: 10
|
| 20 |
+
eval_guidance_scales:
|
| 21 |
+
- 1.0
|
| 22 |
+
- 2.0
|
| 23 |
+
- 3.0
|
| 24 |
+
- 4.5
|
| 25 |
+
eval_prompt_fraction: 1.0
|
| 26 |
+
fd_allow_small_reference: false
|
| 27 |
+
fd_batch_size: 12
|
| 28 |
+
fd_cache_dir: models/fd_stats
|
| 29 |
+
fd_clamp: 10.0
|
| 30 |
+
fd_cov_source: residual
|
| 31 |
+
fd_cov_weight: 1.0
|
| 32 |
+
fd_eig_floor: 0.0
|
| 33 |
+
fd_ema_beta: 0.95
|
| 34 |
+
fd_encoder_cache_dir: models/fd_encoders
|
| 35 |
+
fd_init_epochs: 25
|
| 36 |
+
fd_log_every_n_epochs: 1
|
| 37 |
+
fd_log_types:
|
| 38 |
+
- dinov2_vitl14
|
| 39 |
+
- inception_v3
|
| 40 |
+
- mae_vitl16
|
| 41 |
+
fd_mean_weight: 1.0
|
| 42 |
+
fd_norm_eps: null
|
| 43 |
+
fd_normalize: null_ratio
|
| 44 |
+
fd_reference_seed: null
|
| 45 |
+
fd_scale: 0.0
|
| 46 |
+
fd_shrinkage: 0.02
|
| 47 |
+
fd_train_types:
|
| 48 |
+
- convnextv2_base
|
| 49 |
+
- siglip2_so400m
|
| 50 |
+
- clip_vitl14
|
| 51 |
+
fd_train_weights:
|
| 52 |
+
- 1.0
|
| 53 |
+
- 1.0
|
| 54 |
+
- 1.0
|
| 55 |
+
grad_norm_monitor_every_n_epochs: 1
|
| 56 |
+
integral_kl_timestep_distribution: uniform
|
| 57 |
+
integral_kl_weight: 0.0
|
| 58 |
+
lora_alpha: 64
|
| 59 |
+
lora_rank: 32
|
| 60 |
+
loss_batch_size: 16
|
| 61 |
+
lr: 0.0003
|
| 62 |
+
min_timesteps: 400
|
| 63 |
+
mixed_precision: fp16
|
| 64 |
+
model_name: stabilityai/stable-diffusion-3.5-medium
|
| 65 |
+
num_eval_inference_steps: 40
|
| 66 |
+
num_integral_kl_timesteps: 0
|
| 67 |
+
num_loss_targets_per_sample: 8
|
| 68 |
+
num_optim_steps_per_epoch: 1
|
| 69 |
+
num_prompts_per_epoch: 48
|
| 70 |
+
num_samples_per_prompt: 24
|
| 71 |
+
num_train_inference_steps: 20
|
| 72 |
+
prompts: geneval
|
| 73 |
+
resolution: 512
|
| 74 |
+
reverse_kl_weight: 1.0
|
| 75 |
+
reward_batch_size: 12
|
| 76 |
+
reward_multiplier: 100
|
| 77 |
+
rewards:
|
| 78 |
+
- Geneval
|
| 79 |
+
sampling_batch_size: 32
|
| 80 |
+
scale_rewards: epoch
|
| 81 |
+
seed: 0
|
| 82 |
+
sft_loss_weight: 0.0
|
| 83 |
+
sft_loss_weight_fn: null
|
| 84 |
+
sft_num_loss_targets_per_sample: 8
|
| 85 |
+
sft_num_samples_per_prompt: 24
|
| 86 |
+
sft_ref_cache_dir: null
|
| 87 |
+
timestep_sampling: weighted
|
| 88 |
+
train_guidance_scale: 2.0
|
| 89 |
+
validation_rewards:
|
| 90 |
+
- ImageReward
|
| 91 |
+
wandb_entity: jaylee2000-kaist
|
| 92 |
+
wandb_project: ram
|
| 93 |
+
wandb_run_name: null
|
| 94 |
+
weight_decay: 0.0001
|
geneval_sd3_fd3_kl_phi015/config.yaml
ADDED
|
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
beta1: 0.9
|
| 2 |
+
beta2: 0.99
|
| 3 |
+
dreamsim_batch_size: 12
|
| 4 |
+
dreamsim_cache_dir: models/dreamsim
|
| 5 |
+
dreamsim_eval_guidance_scale: 2.0
|
| 6 |
+
dreamsim_eval_num_prompts: 32
|
| 7 |
+
dreamsim_eval_samples_per_prompt: 8
|
| 8 |
+
dreamsim_gate: std
|
| 9 |
+
dreamsim_log_every_n_epochs: 1
|
| 10 |
+
dreamsim_log_types:
|
| 11 |
+
- dino_vitb16_patch
|
| 12 |
+
- clip_vitb32
|
| 13 |
+
- open_clip_vitb32
|
| 14 |
+
- ensemble
|
| 15 |
+
dreamsim_scale: 0.0
|
| 16 |
+
dreamsim_train_type: dino_vitb16_patch
|
| 17 |
+
ema_decay: 0.9
|
| 18 |
+
ema_warmup_rate: 0.01
|
| 19 |
+
eval_every_n_epochs: 10
|
| 20 |
+
eval_guidance_scales:
|
| 21 |
+
- 1.0
|
| 22 |
+
- 2.0
|
| 23 |
+
- 3.0
|
| 24 |
+
- 4.5
|
| 25 |
+
eval_prompt_fraction: 1.0
|
| 26 |
+
fd_allow_small_reference: false
|
| 27 |
+
fd_batch_size: 12
|
| 28 |
+
fd_cache_dir: models/fd_stats
|
| 29 |
+
fd_clamp: 10.0
|
| 30 |
+
fd_cov_source: residual
|
| 31 |
+
fd_cov_weight: 1.0
|
| 32 |
+
fd_eig_floor: 0.0
|
| 33 |
+
fd_ema_beta: 0.95
|
| 34 |
+
fd_encoder_cache_dir: models/fd_encoders
|
| 35 |
+
fd_init_epochs: 25
|
| 36 |
+
fd_log_every_n_epochs: 1
|
| 37 |
+
fd_log_types:
|
| 38 |
+
- dinov2_vitl14
|
| 39 |
+
- inception_v3
|
| 40 |
+
- mae_vitl16
|
| 41 |
+
fd_mean_weight: 1.0
|
| 42 |
+
fd_norm_eps: null
|
| 43 |
+
fd_normalize: null_ratio
|
| 44 |
+
fd_reference_seed: null
|
| 45 |
+
fd_scale: 0.06939
|
| 46 |
+
fd_shrinkage: 0.02
|
| 47 |
+
fd_train_types:
|
| 48 |
+
- convnextv2_base
|
| 49 |
+
- siglip2_so400m
|
| 50 |
+
- clip_vitl14
|
| 51 |
+
fd_train_weights:
|
| 52 |
+
- 1.412
|
| 53 |
+
- 0.662
|
| 54 |
+
- 1.278
|
| 55 |
+
grad_norm_monitor_every_n_epochs: 1
|
| 56 |
+
integral_kl_timestep_distribution: uniform
|
| 57 |
+
integral_kl_weight: 0.0
|
| 58 |
+
lora_alpha: 64
|
| 59 |
+
lora_rank: 32
|
| 60 |
+
loss_batch_size: 16
|
| 61 |
+
lr: 0.0003
|
| 62 |
+
min_timesteps: 400
|
| 63 |
+
mixed_precision: fp16
|
| 64 |
+
model_name: stabilityai/stable-diffusion-3.5-medium
|
| 65 |
+
num_eval_inference_steps: 40
|
| 66 |
+
num_integral_kl_timesteps: 0
|
| 67 |
+
num_loss_targets_per_sample: 8
|
| 68 |
+
num_optim_steps_per_epoch: 1
|
| 69 |
+
num_prompts_per_epoch: 48
|
| 70 |
+
num_samples_per_prompt: 24
|
| 71 |
+
num_train_inference_steps: 20
|
| 72 |
+
prompts: geneval
|
| 73 |
+
resolution: 512
|
| 74 |
+
reverse_kl_weight: 1.0
|
| 75 |
+
reward_batch_size: 12
|
| 76 |
+
reward_multiplier: 100
|
| 77 |
+
rewards:
|
| 78 |
+
- Geneval
|
| 79 |
+
sampling_batch_size: 32
|
| 80 |
+
scale_rewards: epoch
|
| 81 |
+
seed: 0
|
| 82 |
+
sft_loss_weight: 0.0
|
| 83 |
+
sft_loss_weight_fn: null
|
| 84 |
+
sft_num_loss_targets_per_sample: 8
|
| 85 |
+
sft_num_samples_per_prompt: 24
|
| 86 |
+
sft_ref_cache_dir: null
|
| 87 |
+
timestep_sampling: weighted
|
| 88 |
+
train_guidance_scale: 2.0
|
| 89 |
+
validation_rewards:
|
| 90 |
+
- ImageReward
|
| 91 |
+
wandb_entity: jaylee2000-kaist
|
| 92 |
+
wandb_project: ram
|
| 93 |
+
wandb_run_name: null
|
| 94 |
+
weight_decay: 0.0001
|
geneval_sd3_fd3_kl_phi030/config.yaml
ADDED
|
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
beta1: 0.9
|
| 2 |
+
beta2: 0.99
|
| 3 |
+
dreamsim_batch_size: 12
|
| 4 |
+
dreamsim_cache_dir: models/dreamsim
|
| 5 |
+
dreamsim_eval_guidance_scale: 2.0
|
| 6 |
+
dreamsim_eval_num_prompts: 32
|
| 7 |
+
dreamsim_eval_samples_per_prompt: 8
|
| 8 |
+
dreamsim_gate: std
|
| 9 |
+
dreamsim_log_every_n_epochs: 1
|
| 10 |
+
dreamsim_log_types:
|
| 11 |
+
- dino_vitb16_patch
|
| 12 |
+
- clip_vitb32
|
| 13 |
+
- open_clip_vitb32
|
| 14 |
+
- ensemble
|
| 15 |
+
dreamsim_scale: 0.0
|
| 16 |
+
dreamsim_train_type: dino_vitb16_patch
|
| 17 |
+
ema_decay: 0.9
|
| 18 |
+
ema_warmup_rate: 0.01
|
| 19 |
+
eval_every_n_epochs: 10
|
| 20 |
+
eval_guidance_scales:
|
| 21 |
+
- 1.0
|
| 22 |
+
- 2.0
|
| 23 |
+
- 3.0
|
| 24 |
+
- 4.5
|
| 25 |
+
eval_prompt_fraction: 1.0
|
| 26 |
+
fd_allow_small_reference: false
|
| 27 |
+
fd_batch_size: 12
|
| 28 |
+
fd_cache_dir: models/fd_stats
|
| 29 |
+
fd_clamp: 10.0
|
| 30 |
+
fd_cov_source: residual
|
| 31 |
+
fd_cov_weight: 1.0
|
| 32 |
+
fd_eig_floor: 0.0
|
| 33 |
+
fd_ema_beta: 0.95
|
| 34 |
+
fd_encoder_cache_dir: models/fd_encoders
|
| 35 |
+
fd_init_epochs: 25
|
| 36 |
+
fd_log_every_n_epochs: 1
|
| 37 |
+
fd_log_types:
|
| 38 |
+
- dinov2_vitl14
|
| 39 |
+
- inception_v3
|
| 40 |
+
- mae_vitl16
|
| 41 |
+
fd_mean_weight: 1.0
|
| 42 |
+
fd_norm_eps: null
|
| 43 |
+
fd_normalize: null_ratio
|
| 44 |
+
fd_reference_seed: null
|
| 45 |
+
fd_scale: 0.1388
|
| 46 |
+
fd_shrinkage: 0.02
|
| 47 |
+
fd_train_types:
|
| 48 |
+
- convnextv2_base
|
| 49 |
+
- siglip2_so400m
|
| 50 |
+
- clip_vitl14
|
| 51 |
+
fd_train_weights:
|
| 52 |
+
- 1.412
|
| 53 |
+
- 0.662
|
| 54 |
+
- 1.278
|
| 55 |
+
grad_norm_monitor_every_n_epochs: 1
|
| 56 |
+
integral_kl_timestep_distribution: uniform
|
| 57 |
+
integral_kl_weight: 0.0
|
| 58 |
+
lora_alpha: 64
|
| 59 |
+
lora_rank: 32
|
| 60 |
+
loss_batch_size: 16
|
| 61 |
+
lr: 0.0003
|
| 62 |
+
min_timesteps: 400
|
| 63 |
+
mixed_precision: fp16
|
| 64 |
+
model_name: stabilityai/stable-diffusion-3.5-medium
|
| 65 |
+
num_eval_inference_steps: 40
|
| 66 |
+
num_integral_kl_timesteps: 0
|
| 67 |
+
num_loss_targets_per_sample: 8
|
| 68 |
+
num_optim_steps_per_epoch: 1
|
| 69 |
+
num_prompts_per_epoch: 48
|
| 70 |
+
num_samples_per_prompt: 24
|
| 71 |
+
num_train_inference_steps: 20
|
| 72 |
+
prompts: geneval
|
| 73 |
+
resolution: 512
|
| 74 |
+
reverse_kl_weight: 1.0
|
| 75 |
+
reward_batch_size: 12
|
| 76 |
+
reward_multiplier: 100
|
| 77 |
+
rewards:
|
| 78 |
+
- Geneval
|
| 79 |
+
sampling_batch_size: 32
|
| 80 |
+
scale_rewards: epoch
|
| 81 |
+
seed: 0
|
| 82 |
+
sft_loss_weight: 0.0
|
| 83 |
+
sft_loss_weight_fn: null
|
| 84 |
+
sft_num_loss_targets_per_sample: 8
|
| 85 |
+
sft_num_samples_per_prompt: 24
|
| 86 |
+
sft_ref_cache_dir: null
|
| 87 |
+
timestep_sampling: weighted
|
| 88 |
+
train_guidance_scale: 2.0
|
| 89 |
+
validation_rewards:
|
| 90 |
+
- ImageReward
|
| 91 |
+
wandb_entity: jaylee2000-kaist
|
| 92 |
+
wandb_project: ram
|
| 93 |
+
wandb_run_name: null
|
| 94 |
+
weight_decay: 0.0001
|
geneval_sd3_fd3_phi015/step_10/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8b3cad28e236ac623f858d6b95798ca494c6629aec7c07f05f367089258f7912
|
| 3 |
+
size 5617919800
|
geneval_sd3_fd3_phi015/step_20/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:92afe3b2e87579ea2cd0b1a8f2fe940ac199f8d2a661a7ce3dcd58b1df4c4f07
|
| 3 |
+
size 5617919800
|
geneval_sd3_fd3_phi015/step_30/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:77c3f0d786da52089f8fbce70718eea84ecc1a20e5f8049241ea51b54291947e
|
| 3 |
+
size 5617919800
|
geneval_sd3_fd3_phi015/step_40/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f454358e965d43afec7feb0f21cbf108880ae505b2cbf965536c0ac822d84009
|
| 3 |
+
size 5617919800
|