reference/methods.md

# Training and evaluation methods

[Contents](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/README.md)

## Training an assistant and a concept

Train the assistant adapter on ordinary image targets. Then freeze it at strength 1, add a fresh concept adapter, and train the concept adapter on the downstream data. For concept validation and inference, disable the assistant and retain the learned concept adapter.

The working hypothesis is that a removable assistant can carry changes associated with losing the base model's distilled behaviour. Compare downstream concept learning, image coherence, and changes to unrelated subjects separately.

## Shared concept-training settings

Unless a chapter specifies otherwise: 27 Domokun images with caption `🟫`, rank/alpha 32/32, batch/accumulation 1/1, training seed 42, BF16, corrected `adamw_bf16`, norm clipping 1.0, and constant learning rate after 25 warmup updates. The base revision is `790c92633540aa0cb11d9abf19eb46d861714758`. VAE encoding and decoding use full frames.

## Evaluation settings

| Protocol | Generator / seed | CFG | Output |
| --- | --- | --- | --- |
| Periodic validation | CPU / 42 | 1 | 512px or 1024px, labelled with each grid |
| Independent inference | CUDA / 123 or 777, as labelled | 1 or 4, as labelled | 512px |

All comparisons use 40 inference steps and the Details-fixed VAE (madebyollin). Independent inference uses adapter strength 1. True CFG 4 uses conditional and unconditional passes with an empty negative prompt: 80 audited transformer forwards, versus 40 at CFG 1. Generator device is part of the seed specification.

## Reading the grids

Columns show prompts and rows show checkpoints or training methods. Resolution, CFG, and seed are separated into labelled grids. Click an image for full resolution. Missing source samples remain empty rather than being substituted from another setting.

Periodic panels are lossless crops of saved validation grids. The [panel manifest](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/panels.json) records their source grid, crop coordinates, and hashes; the original files remain available. Base panels are shared where their pixels are identical.

[Exact prompts](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/prompts.json) · [Image provenance](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/images.json) · [Adapter hashes](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/artifacts.json) · [Training provenance](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/training_provenance.json)

## Limitations

Each training variant has one training seed. Historical runs differ in source-image processing, cache history, prompt sets, and some training settings. The seed-123/777 screen followed prompt exploration. Observations are qualitative; they do not establish a general preservation guarantee or isolate loss of CFG distillation as the cause of degradation.

[Contents](https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments/blob/main/README.md)