no safe dose
Collection
33 items • Updated
Dose-Response C1 (8M-0%) with SafeCLIP: All unsafe removed
Text-encoder ablation using SafeCLIP (77 tokens) in place of the default T5-Gemma-2B (256 tokens).
Condition
| Label | C1 (8M-0%) |
| Description | All unsafe images removed (N kept approximately fixed). |
| Training set size N | 7.94M |
| Unsafe fraction p | 0% |
| Unsafe count U | 0 |
Architecture
| Class | PRX (rectified-flow DiT) |
| Hidden size | 1792 |
| Depth | 16 |
| Heads | 28 |
| MLP ratio | 3.5 |
| Patch size | 32 px |
| Bottleneck | 256 |
| Resolution | 512×512 |
Text encoder
| Model | aimagelab/safeclip_vit-l_14 |
| Max prompt tokens | 77 |
| Dtype | bfloat16 |
Diffusion scheduler
| Type | x-prediction flow matching |
| Train timesteps | 1000 |
| Timestep shift | 3.0 |
Training
| Iterations | 100,000 |
| Samples seen | ~25.60M |
| Global batch size | 256 |
| Microbatch (per GPU) | 32 |
| Hardware | 8× NVIDIA H200 |
| Precision | bfloat16 (amp_bf16) |
| Optimizer (transformer blocks) | Muon (lr=1e-4, momentum=0.95, nesterov, ns_steps=5, weight_decay=0) |
| Optimizer (other params) | AdamW (lr=1e-4, β=(0.9, 0.95), eps=1e-8, weight_decay=0) |
| LR schedule | 1,000-step linear warmup, constant after |
| EMA | decay 0.999, started at step 0 |
| Random seed | 42 |
| Trainer | Composer + FSDP |
Training data sources
The training set combines three image datasets, with per-condition filtering/oversampling:
lehduong/flux_generated(~1.7M)LucasFang/FLUX-Reason-6M(~6M)brivangl/midjourney-v6-llava(~1M)
Files
denoiser.pt— Consolidated EMA-denoiser checkpointconfig.yaml— Full training configuration
Framework
Trained with the PRX framework (Composer + FSDP). The full config.yaml is included for reproducibility.
- Downloads last month
- 18