Fyre, a fictional character, redhead with freckles.Trained on 20 Klein 9b generated images generated from an initial z-image generated image.Trained with diffusion-pipe, this is a 128-rank lora trained on z-image (non-turbo, aka "base"), wan 2.2 14B low noise, and qwen image 2512, using the ProdigyPlusScheduleFree optimizer (<a target="_blank" rel="ugc" href="https://github.com/tdrussell/diffusion-pipe/pull/483">https://github.com/tdrussell/diffusion-pipe/pull/483</a>), and masked training with backgrounds masked to 10%.The model was evaluated every 200 steps (10 repeats of each training image) using FaceNet512 to evaluate the face similarity, using the same settings, seed, etc for inference, calculating the average distance to a subset of 10 of the training images. For the z-image model, the best face similarity was achieved around epoch 16 and 17 (3,200 steps and 3,400 steps). For the qwen image 2512 model, the best face similarity was achieved at epoch 20 (4,000 steps). For wan 2.2 low noise, the best face similarity was achieved at epoch 10 (2,000) steps.The attached training data zip includes: the diffusion pipe config files; the training images, prompts, and masks; the validation images (subset of training images); and - for the z-image model - the output of <a target="_blank" rel="ugc" href="https://civarchive.com/articles/26241/stop-training-on-vibes-a-visual-guide-to-biometric-lora-diagnosis-mirror-metrics">MirrorMetrics</a> evaluation of epoch16 and 17 vs. the validation set.z-image model evaluation:<ul><li>For average distances, lower is better. Distance under 0.3 is considered a face match.</li><li>Matching percentage is the percent of the 10 validation images that the generated face was under 0.3 distance from.</li></ul><ul><li>epoch 16 (3,200 steps)<ul><li>z-image (non-turbo) inference<ul><li>Average distance: 0.187</li><li>Matching percentage: 100%</li></ul></li><li>z-image-turbo inference<ul><li>Average distance: 0.262</li><li>Matching percentage: 80%</li></ul></li></ul></li></ul><ul><li>epoch 17 (3,400 steps)<ul><li>z-image (non-turbo) inference<ul><li>Average distance: 0.193</li><li>Matching percentage: 100%</li></ul></li><li>z-image-turbo inference<ul><li>Average distance: 0.255</li><li>Matching percentage: 80%</li></ul></li></ul></li></ul>Evaluation using <a target="_blank" rel="ugc" href="https://civarchive.com/articles/26241/stop-training-on-vibes-a-visual-guide-to-biometric-lora-diagnosis-mirror-metrics">MirrorMetrics</a> (html output included in training data attachment)<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/5c318714-2dee-4c9f-ab2e-e6161b646b8d/original=true/5c318714-2dee-4c9f-ab2e-e6161b646b8d.jpeg" />Training config:(included in attached training data attachment)Training command:<pre><code>PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True NCCL_P2P_DISABLE="1" NCCL_IB_DISABLE="1" deepspeed \
 --num_gpus=1 train.py \ 
 --deepspeed \ 
 --regenerate_cache \
 --config /home/user/diffusion-pipe-config.toml</code></pre>

Fyre (fictional character)

Wan 2.2 14B low noise, 10 epochs, 2,000 steps

Files with SHA256: fa7bd50d4852da85e82a591e2d25deb6d9ce708ad2949f26b345d857ede0ccce