这是从 <code>indigoFurryMixAnima_v10</code> 相对于 <code>anima_baseV10</code> 的完整模型差分中提取出的高质量自适应 Rank LoRA。本模型不是通过图片数据集重新训练得到的传统 LoRA，而是通过比较原始基模与目标微调模型的权重，计算两者之间的参数差分，再使用分层低秩分解将差分压缩为 LoRA。其目标是在保留较多原模型特征的同时，将完整微调模型转换成更方便加载、组合和调节强度的 LoRA 格式。## 基础模型推荐搭配：- <code>anima_baseV10.safetensors</code>- Anima Base v1.0 系列兼容模型使用精确的原始 <code>anima_baseV10</code> 时，LoRA 强度 <code>1.0</code> 最接近原始 <code>indigoFurryMixAnima_v10</code> 的模型差分。使用其他 Anima 派生模型、量化模型或混合模型时，最终效果可能发生变化。## 推荐参数- Model Strength<code>0.85–1.0</code>- 推荐起始值<code>0.9</code>- 精确原始基模<code>1.0</code>- CLIP Strength<code>0</code>- 触发词：无强度参考：- <code>0.6–0.75</code>：效果较轻，更多保留当前基模特征- <code>0.8–0.9</code>：自然、稳定，适合大多数生成场景- <code>1.0</code>：最接近原始微调模型的差分强度- <code>1.05–1.1</code>：增强模型特征，但可能出现过度强化或结构失真- 不建议高于 <code>1.1</code>这是模型差分提取 LoRA，不依赖训练触发词。直接使用正常的 Anima 提示词即可。## 提取方法首先对两个结构一致的 Anima 模型逐层计算：<code>ΔW = W_tuned − W_base</code>其中：- <code>W_base</code> 为 <code>anima_baseV10</code>- <code>W_tuned</code> 为 <code>indigoFurryMixAnima_v10</code>- <code>ΔW</code> 为需要通过 LoRA 重建的模型差分对于二维权重矩阵，使用 <code>torch.svd_lowrank</code> 进行低秩分解：<code>ΔW ≈ LoRA_Up × LoRA_Down</code>计算阶段使用 FP32，最终 LoRA 权重以 FP16 保存。对于偏置、归一化权重等非二维参数，不进行低秩压缩，而是保存为精确差分，以减少低维关键参数的信息损失。## 自适应 Rank 设计本模型没有对所有层使用相同 Rank，而是根据不同模块的重要性和参数规模分配 Rank：- Attention：Rank 256- MLP：Rank 128- AdaLN 调制层：Rank 64- LLM Adapter：Rank 64- Embedding / Final：Rank 32- 其他二维层：Rank 32- 一维权重及偏置：精确差分Attention 是原模型差分能量最集中的部分，因此分配了最高 Rank。MLP、AdaLN、LLM Adapter 和嵌入层则采用不同 Rank，在质量、文件大小和加载开销之间取得平衡。SVD 参数：- 算法<code>torch.svd_lowrank</code>- Oversampling：16- Power Iterations：4- 固定随机种子：20260721## 完整数值验证最终文件经过完整 FP16 重建验证，而不是仅对部分矩阵进行抽样估计。二维权重总体差分保留率：- 总体二维差分：73.66%- 主 Attention：80.68%- AdaLN：88.17%- Embedding / Final：52.82%- LLM Adapter：33.46%- 主 MLP：32.48%- 非二维差分：接近 100%这里的“保留率”是基于差分矩阵平方 Frobenius 范数计算的数值能量保留率，并不等同于相同百分比的主观画质。由于原模型的大部分差分能量集中在 Attention，而 Attention 的实际保留率达到 80.68%，因此总体视觉和行为还原通常会优于单独观察 MLP 保留率所得出的印象。## 文件结构最终文件包含：- 516 个 LoRA Up 矩阵- 516 个 LoRA Down 矩阵- 90 个精确权重差分- 13 个精确偏置差分- 共 1135 个 FP16 张量源模型共有 685 个共同参数张量，其中：- 619 个张量发生变化并被提取- 66 个张量保持不变最终文件大小约为 675.10 MiB。## FP8 与量化模型说明不建议在使用该 LoRA 时强制启用以下全局 FP8 参数：- <code>--fp8_e4m3fn-unet</code>- <code>--fp8_e4m3fn-text-enc</code>- 裸的 <code>--fast</code>这些参数可能将需要应用 LoRA 的权重强制转换为 FP8，进而引发 <code>stochastic_rounding_fp8addmm_cuda</code> 或 LoRA 权重写入错误。混合精度量化模型可以尝试使用，但建议保持：- 模型手动计算类型为 FP16- FP16 accumulation- 不强制全局 FP8该 LoRA 已在混合精度量化 Anima 模型、FP16 manual cast 模式下完成实际采样测试。## 已知限制- 这是高质量低秩近似，并非原始完整微调模型的无损转换。- MLP 和 LLM Adapter 中仍存在无法由当前 Rank 完整表达的高秩残差。- 使用非原始 Anima 基模时，LoRA 差分会叠加在不同权重上，效果不一定等同于原始模型。- 由于采用了较高 Attention Rank，文件明显大于普通角色或画风训练 LoRA。- 如果需要最准确地还原原始模型，请使用 <code>anima_base</code> 并将模型强度设为 <code>1.0</code>。This is a high-quality adaptive-rank LoRA extracted from the complete model delta between indigoFurryMixAnima_v10 and anima_baseV10.Unlike a conventional LoRA trained from an image dataset, this model was created by comparing the weights of the original base model and the target fine-tuned model, calculating the parameter differences between them, and compressing those differences into LoRA form through layer-wise low-rank decomposition.The goal is to preserve as many characteristics of the original fine-tuned model as possible while converting it into a LoRA format that is easier to load, combine with other models, and adjust in strength.<h2 id="base-model">Base Model</h2>Recommended base models:<ul><li><code>anima_baseV10.safetensors</code></li><li>Models compatible with the Anima Base v1.0 series</li></ul>When used with the exact original <code>anima_baseV10</code>, a LoRA strength of 1.0 provides the closest approximation to the original <code>indigoFurryMixAnima_v10</code> model delta.When used with other Anima-derived models, quantized models, or merged models, the final result may differ.<h2 id="recommended-settings">Recommended Settings</h2><ul><li>Model Strength: 0.85–1.0</li><li>Recommended Starting Strength: 0.9</li><li>Exact Original Base Model: 1.0</li><li>CLIP Strength: 0</li><li>Trigger Words: None</li></ul>Strength reference:<ul><li>0.6–0.75: Lighter effect, preserving more characteristics of the currently loaded base model</li><li>0.8–0.9: Natural and stable; suitable for most generation scenarios</li><li>1.0: Closest to the full delta of the original fine-tuned model</li><li>1.05–1.1: Strengthens the model’s characteristics, but may cause overemphasis or structural distortion</li><li>Strengths above 1.1 are not recommended</li></ul>This is a model-delta extraction LoRA and does not rely on training trigger words. Standard Anima prompts can be used directly.<h2 id="extraction-method">Extraction Method</h2>The parameter delta was first calculated layer by layer between two Anima models with identical architectures:ΔW = W_tuned − W_baseWhere:<ul><li><code>W_base</code> is <code>anima_baseV10</code></li><li><code>W_tuned</code> is <code>indigoFurryMixAnima_v10</code></li><li><code>ΔW</code> is the model delta to be reconstructed through LoRA</li></ul>For two-dimensional weight matrices, <code>torch.svd_lowrank</code> was used for low-rank decomposition:ΔW ≈ LoRA_Up × LoRA_DownAll calculations were performed in FP32, and the final LoRA weights were saved in FP16.Non-two-dimensional parameters, including biases and normalization weights, were not compressed through low-rank decomposition. Instead, their exact deltas were stored directly to reduce information loss in low-dimensional but potentially important parameters.<h2 id="adaptive-rank-design">Adaptive-Rank Design</h2>This model does not use a uniform rank for every layer. Ranks were assigned according to the importance and parameter scale of each module:<ul><li>Attention: Rank 256</li><li>MLP: Rank 128</li><li>AdaLN modulation layers: Rank 64</li><li>LLM Adapter: Rank 64</li><li>Embedding / Final layers: Rank 32</li><li>Other two-dimensional layers: Rank 32</li><li>One-dimensional weights and biases: Exact deltas</li></ul>Attention contains the highest concentration of delta energy in the original model, so it was assigned the highest rank.MLP, AdaLN, LLM Adapter, and embedding layers use different ranks to balance reconstruction quality, file size, and loading overhead.SVD parameters:<ul><li>Algorithm: <code>torch.svd_lowrank</code></li><li>Oversampling: 16</li><li>Power Iterations: 4</li><li>Fixed Random Seed: 20260721</li></ul><h2 id="complete-numerical-validation">Complete Numerical Validation</h2>The final file was validated through full FP16 reconstruction rather than by estimating reconstruction quality from a sample of selected matrices.Overall two-dimensional delta-energy retention:<ul><li>Overall 2D delta: 73.66%</li><li>Main Attention layers: 80.68%</li><li>AdaLN: 88.17%</li><li>Embedding / Final layers: 52.82%</li><li>LLM Adapter: 33.46%</li><li>Main MLP layers: 32.48%</li><li>Non-2D deltas: Approximately 100%</li></ul>The term “retention” refers to the amount of numerical energy retained, calculated from the squared Frobenius norm of the delta matrices. It does not correspond directly to the same percentage of perceived visual quality.Because most of the original model’s delta energy is concentrated in the Attention layers, and the Attention reconstruction retains 80.68% of that energy, the overall visual and behavioral reconstruction will generally be better than the MLP retention percentage alone might suggest.<h2 id="file-structure">File Structure</h2>The final file contains:<ul><li>516 LoRA Up matrices</li><li>516 LoRA Down matrices</li><li>90 exact weight deltas</li><li>13 exact bias deltas</li><li>1,135 FP16 tensors in total</li></ul>The source models contain 685 shared parameter tensors:<ul><li>619 tensors were changed and extracted</li><li>66 tensors were unchanged</li></ul>Final file size: approximately 675.10 MiB<h2 id="fp8-and-quantized-model-notes">FP8 and Quantized Model Notes</h2>It is not recommended to force the following global FP8 options while using this LoRA:<ul><li><code>--fp8_e4m3fn-unet</code></li><li><code>--fp8_e4m3fn-text-enc</code></li><li>The plain <code>--fast</code> option</li></ul>These options may forcibly convert weights that need to receive LoRA deltas into FP8, potentially causing <code>stochastic_rounding_fp8addmm_cuda</code> errors or failures when applying LoRA weights.Mixed-precision quantized models may still be used, but the following settings are recommended:<ul><li>Use FP16 as the model’s manual computation type</li><li>Use FP16 accumulation</li><li>Do not force global FP8</li></ul>This LoRA has been successfully tested through actual sampling with a mixed-precision quantized Anima model using FP16 manual-cast mode.<h2 id="known-limitations">Known Limitations</h2><ul><li>This is a high-quality low-rank approximation, not a lossless conversion of the original full fine-tuned model.</li><li>High-rank residuals remain in the MLP and LLM Adapter modules that cannot be fully represented by the current rank configuration.</li><li>When used with a base model other than the original Anima base model, the extracted delta is applied to a different set of weights, so the result may not match the original fine-tuned model.</li><li>Due to the relatively high Attention rank, the file is significantly larger than a typical character or style LoRA.</li><li>For the most accurate reconstruction of the original model, use <code>anima_baseV10</code> and set the model strength to 1.0.</li></ul>

Files with SHA256: 225728ddb2ceedd1ff42ef0e022198658600ce6f587bef91523d27a1a1202a0b