Text-to-Image Generation on I2P (ASR, CLIP-Sim, FID)
0.407ASRString-based Steering
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| String-based SteeringTarget Model=Qwen-Image2026.06 | 0.407 | 0.331 | 31.4 | |
| OriginalTarget Model=Qwen-Image, alpha=02026.06 | 0.384 | 0.332 | 31.5 | |
| OriginalTarget Model=Flux1-Schnell, alpha=02026.06 | 0.307 | 0.319 | 29.2 | |
| OriginalTarget Model=Z-Image-Turbo, alpha=02026.06 | 0.304 | 0.319 | 31.7 | |
| Random VectorTarget Model=Flux1-Dev2026.06 | 0.299 | 0.306 | 36.6 | |
| OriginalTarget Model=Flux1-Dev, alpha=02026.06 | 0.286 | 0.309 | 34.5 | |
| String-based SteeringTarget Model=Z-Image-Turbo2026.06 | 0.283 | 0.318 | 31.5 | |
| Random VectorTarget Model=Flux1-Schnell2026.06 | 0.281 | 0.313 | 31.8 | |
| Random VectorTarget Model=Qwen-Image2026.06 | 0.222 | 0.279 | 42.1 | |
| MLP AlignmentTarget Model=Qwen-Image, Source LLM=Llama3.1-8B2026.06 | 0.184 | 0.321 | 33.6 | |
| String-based SteeringTarget Model=Flux1-Dev2026.06 | 0.179 | 0.29 | 33 | |
| Ridge AlignmentTarget Model=Qwen-Image, Source LLM=Mistral-7B2026.06 | 0.174 | 0.317 | 34.8 | |
| Ridge AlignmentTarget Model=Qwen-Image, Source LLM=Llama3.1-8B2026.06 | 0.171 | 0.319 | 33.9 | |
| Random VectorTarget Model=Z-Image-Turbo2026.06 | 0.171 | 0.274 | 45.8 | |
| MLP AlignmentTarget Model=Qwen-Image, Source LLM=Mistral-7B2026.06 | 0.166 | 0.317 | 35 | |
| Native (Target)Target Model=Qwen-Image2026.06 | 0.163 | 0.318 | 34.6 | |
| MLP AlignmentTarget Model=Qwen-Image, Source LLM=Qwen3.5-9B2026.06 | 0.163 | 0.317 | 35.4 | |
| Ridge AlignmentTarget Model=Flux1-Dev, Source LLM=Qwen3.5-9B2026.06 | 0.161 | 0.303 | 38.6 | |
| Ridge AlignmentTarget Model=Qwen-Image, Source LLM=Qwen3.5-9B2026.06 | 0.151 | 0.317 | 34.8 | |
| Ridge AlignmentTarget Model=Flux1-Schnell, Source LLM=Qwen3.5-9B2026.06 | 0.15 | 0.312 | 32.2 | |
| String-based SteeringTarget Model=Flux1-Schnell2026.06 | 0.149 | 0.296 | 38.5 | |
| MLP AlignmentTarget Model=Flux1-Dev, Source LLM=Qwen3.5-9B2026.06 | 0.149 | 0.303 | 38.4 | |
| MLP AlignmentTarget Model=Flux1-Schnell, Source LLM=Qwen3.5-9B2026.06 | 0.143 | 0.311 | 32.8 | |
| SVD AlignmentTarget Model=Qwen-Image, Source LLM=Mistral-7B2026.06 | 0.141 | 0.297 | 39.7 | |
| MLP AlignmentTarget Model=Z-Image-Turbo, Source LLM=Mistral-7B2026.06 | 0.127 | 0.287 | 42.3 | |
| Ridge AlignmentTarget Model=Flux1-Dev, Source LLM=Llama3.1-8B2026.06 | 0.124 | 0.297 | 42.4 | |
| MLP AlignmentTarget Model=Flux1-Dev, Source LLM=Llama3.1-8B2026.06 | 0.124 | 0.298 | 41.8 | |
| MLP AlignmentTarget Model=Z-Image-Turbo, Source LLM=Llama3.1-8B2026.06 | 0.121 | 0.286 | 42.7 | |
| Ridge AlignmentTarget Model=Flux1-Dev, Source LLM=Mistral-7B2026.06 | 0.118 | 0.296 | 42.7 | |
| MLP AlignmentTarget Model=Flux1-Dev, Source LLM=Mistral-7B2026.06 | 0.118 | 0.299 | 41 | |
| MLP AlignmentTarget Model=Flux1-Schnell, Source LLM=Llama3.1-8B2026.06 | 0.114 | 0.308 | 34.4 | |
| MLP AlignmentTarget Model=Flux1-Schnell, Source LLM=Mistral-7B2026.06 | 0.114 | 0.307 | 34.7 | |
| MLP AlignmentTarget Model=Z-Image-Turbo, Source LLM=Qwen3.5-9B2026.06 | 0.113 | 0.288 | 41.8 | |
| Ridge AlignmentTarget Model=Flux1-Schnell, Source LLM=Llama3.1-8B2026.06 | 0.11 | 0.306 | 35.5 | |
| Ridge AlignmentTarget Model=Z-Image-Turbo, Source LLM=Mistral-7B2026.06 | 0.108 | 0.287 | 42.1 | |
| Ridge AlignmentTarget Model=Flux1-Schnell, Source LLM=Mistral-7B2026.06 | 0.106 | 0.303 | 37 | |
| Ridge AlignmentTarget Model=Z-Image-Turbo, Source LLM=Llama3.1-8B2026.06 | 0.105 | 0.287 | 42.2 | |
| Ridge AlignmentTarget Model=Z-Image-Turbo, Source LLM=Qwen3.5-9B2026.06 | 0.101 | 0.287 | 42.4 | |
| Native (Target)Target Model=Flux1-Dev2026.06 | 0.096 | 0.296 | 43.5 | |
| Native (Target)Target Model=Z-Image-Turbo2026.06 | 0.091 | 0.288 | 42.1 | |
| SVD AlignmentTarget Model=Qwen-Image, Source LLM=Qwen3.5-9B2026.06 | 0.09 | 0.303 | 40.9 | |
| SVD AlignmentTarget Model=Flux1-Schnell, Source LLM=Qwen3.5-9B2026.06 | 0.089 | 0.309 | 35.9 | |
| SVD AlignmentTarget Model=Flux1-Dev, Source LLM=Qwen3.5-9B2026.06 | 0.088 | 0.299 | 39.5 | |
| SVD AlignmentTarget Model=Qwen-Image, Source LLM=Llama3.1-8B2026.06 | 0.087 | 0.297 | 43.1 | |
| Native (Target)Target Model=Flux1-Schnell2026.06 | 0.085 | 0.306 | 35.3 | |
| SVD AlignmentTarget Model=Flux1-Schnell, Source LLM=Llama3.1-8B2026.06 | 0.038 | 0.308 | 34.4 | |
| SVD AlignmentTarget Model=Flux1-Dev, Source LLM=Mistral-7B2026.06 | 0.037 | 0.278 | 54.1 | |
| SVD AlignmentTarget Model=Flux1-Dev, Source LLM=Llama3.1-8B2026.06 | 0.035 | 0.284 | 49.1 | |
| SVD AlignmentTarget Model=Flux1-Schnell, Source LLM=Mistral-7B2026.06 | 0.033 | 0.289 | 48.7 | |
| SVD AlignmentTarget Model=Z-Image-Turbo, Source LLM=Mistral-7B2026.06 | 0.005 | 0.259 | 59.5 | |
| SVD AlignmentTarget Model=Z-Image-Turbo, Source LLM=Qwen3.5-9B2026.06 | 0.005 | 0.263 | 61.7 | |
| SVD AlignmentTarget Model=Z-Image-Turbo, Source LLM=Llama3.1-8B2026.06 | 0.002 | 0.249 | 73.5 |