# Combining an assistant with regularisation [Contents](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/README.md) · [Previous: Synthetic regularisation](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/experiments/synthetic-regularisation.md) · [Next: Multi-scale training and REPA](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/experiments/multiscale-repa.md) Combining assistant v2 with synthetic regularisation produces coherent scenes, portraits, and fox illustrations in the inspected final samples. At both output sizes, the 512px-trained adapter expresses the character in the beach, scarf, and bicycle scenes. Body proportions and bicycle interactions remain imperfect. | Setting | Value | | --- | --- | | Training resolution | Separate 512px and 1024px runs | | Assistant | Frozen v2, strength 1; disabled for validation | | Regularisation target | **Base + frozen assistant v2** | | Student | Base + frozen assistant v2 + trainable concept | | Data | 27 Domokun images + 834 synthetic images per resolution | | Learning rate | `1e-4` | | Updates at midpoint | 439 Domokun + 561 regularisation | | Updates at final | 904 Domokun + 1,096 regularisation | The runs share initial concept weights and the backend-choice sequence. GPU audits verified the assisted target and restoration of the frozen-assistant training stack. ## Final checkpoints The 1024px-trained bicycle sample at 1024px output resembles a human in a creature costume. Its scarf colour also changes with output resolution. CPU-generator seed 42 · 40 inference steps · CFG 1 · Details-fixed VAE (madebyollin) · full-frame decoding. Concept adapter only. Columns show prompts; rows show checkpoints or methods. Click a sample to view it at full resolution. ### 512px output
Checkpoint / methodBeachRed scarfBicycleFoxPortrait
BaseBase, 512pxBase, 512pxBase, 512pxBase, 512pxBase, 512px
512px training512px training, 512px512px training, 512px512px training, 512px512px training, 512px512px training, 512px
1024px training1024px training, 512px1024px training, 512px1024px training, 512px1024px training, 512px1024px training, 512px
### 1024px output
Checkpoint / methodBeachRed scarfBicycleFoxPortrait
BaseBase, 1024pxBase, 1024pxBase, 1024pxBase, 1024pxBase, 1024px
512px training512px training, 1024px512px training, 1024px512px training, 1024px512px training, 1024px512px training, 1024px
1024px training1024px training, 1024px1024px training, 1024px1024px training, 1024px1024px training, 1024px1024px training, 1024px
## Training progression The following columns combine training resolution and prompt; every sample is 1024px output. ### 1024px output
Checkpoint / method512px training: beach512px training: bicycle1024px training: beach1024px training: bicycle
BaseBase, 1024pxBase, 1024pxBase, 1024pxBase, 1024px
Update 1,000Update 1,000, 1024pxUpdate 1,000, 1024pxUpdate 1,000, 1024pxUpdate 1,000, 1024px
Update 2,000Update 2,000, 1024pxUpdate 2,000, 1024pxUpdate 2,000, 1024pxUpdate 2,000, 1024px
## Reproduce 512px: [Training configuration](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/recipes/assisted-v2-regularised-512/config.json) · [Datasets](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/recipes/assisted-v2-regularised-512/dataloader.json) · [Prompts](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/recipes/assisted-v2-regularised-512/prompts.json) · [1,000 updates](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/resolve/main/checkpoints/assisted-v2-regularised-512/step-1000/pytorch_lora_weights.safetensors) · [2,000 updates](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/resolve/main/checkpoints/assisted-v2-regularised-512/step-2000/pytorch_lora_weights.safetensors) 1024px: [Training configuration](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/recipes/assisted-v2-regularised-1024/config.json) · [Datasets](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/recipes/assisted-v2-regularised-1024/dataloader.json) · [Prompts](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/recipes/assisted-v2-regularised-1024/prompts.json) · [1,000 updates](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/resolve/main/checkpoints/assisted-v2-regularised-1024/step-1000/pytorch_lora_weights.safetensors) · [2,000 updates](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/resolve/main/checkpoints/assisted-v2-regularised-1024/step-2000/pytorch_lora_weights.safetensors) These recipes require assistant-aware parent capture. Use the recorded [implementation and launch instructions](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/reference/reproduction.md#assistant-aware-regularisation) and [training provenance](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/training_provenance.json). ## Limitations The synthetic pools differ between training resolutions. [Multi-scale and REPA experiments](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/experiments/multiscale-repa.md) use a bare-base regularisation target, a different training setup. [Contents](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/README.md) · [Previous: Synthetic regularisation](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/experiments/synthetic-regularisation.md) · [Next: Multi-scale training and REPA](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/experiments/multiscale-repa.md)