Fyre, a fictional character, redhead with freckles.Trained on 20 Klein 9b generated images generated from an initial z-image generated image.<h3 id="moss-image">moss image</h3>Trained with ai-toolkit, lokr factor 4, automagic v3 LR 0.0001, weight decay 0.0002, Shift/Balanced. Training config is in Other Files as <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, &quot;system-ui&quot;, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:14px">config-moss-v01.yml<h3 id="qwen-image-2.1">qwen image 2.1</h3>Trained with ai-toolkit, lokr factor 4, automagic v3, LR 0.0005. 1,800 steps outperformed 2,000 steps for facial match.ai-toolkit configuration attached as <a target="_blank" rel="ugc" href="http://config-qwen21.toml.zip">config-qwen21.toml.zip</a><h3 id="minimax-h3">minimax h3</h3>Trained with ai-toolkit, automagic v3 with Contrastive Guidance and training adapter v1; differential guidance scale = 1; rank 24. ai-toolkit configuration attached as minimaxh3-fyre-v01.yaml<h3 id="krea-2">krea 2</h3>Trained with ai-toolkit, automagic v3 optimizer, LR/weight decay 0.0001. Training images, captions, and ai-toolkit config are included in <a target="_blank" rel="ugc" href="http://fyre-krea2-v01.zip">fyre-krea2-v01.zip</a><h3 id="ideogram-4">ideogram 4</h3>Trained with ai-toolkit, automagic v2 optimizer, LR 0.0001, rank 32, with captions generated by ai-toolkit (including bounding boxes) with some manual revisions to include the "Fyre" keyword. Training images, captions, and ai-toolkit config are included in <a target="_blank" rel="ugc" href="http://fyre-ideogram-v01.zip">fyre-ideogram-v01.zip</a><h3 id="z-image-qwen-2512-wan-2.2-14b-low-noise">z-image / qwen 2512 / wan 2.2 14b low noise</h3>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)

config-moss-v01.yml

moss-fyre-v01.safetensors

moss-v01

qwen21-fyre-v01

minimaxh3-v01

krea2-v01

ideogram4-v01

wan-2.2-low-noise-v01-e10

qwen-2512-v01-e20

zib-v01-e17

zib-v01-e16

Files with SHA256: 06efbf8fbf18eb21a68484a89e8022b540fb3682f461385c5517c17b14409cad