<h3 id="core-idea">Core Idea</h3><p>Use Anima as a pose-conditioning model to generate initial conditioned latent noise (conditioned on pose information) for the Krea 2 model, which then serves as the starting point for its denoising process.</p><h3 id="variant-1:-krea2_with_anima_sampling-(full-sampling-by-anima-refinement-by-krea-2)">Variant 1: <code>Krea2_with_Anima_Sampling</code> (Full Sampling by Anima, Refinement by Krea 2)</h3><p><strong>Mechanism</strong>: Anima performs the entire denoising sampling process and outputs the final latents. These latents are then passed to Krea 2 for subsequent refinement.</p><p>- <strong>Advantage</strong>: Since Anima dominates the full denoising trajectory, the composition and global structure are highly controllable, resulting in stable and consistent outputs.</p><p>- <strong>Disadvantage</strong>: The composition is fully determined during the Anima phase, leaving little room for character LoRA to influence clothing details during refinement, thus lower fidelity in costume reproduction.</p><h3 id="variant-2:-krea2_with_anima_pre-sampling-(early-stop-by-anima-handover-to-krea-2-for-remaining-denoising)">Variant 2: <code>Krea2_with_Anima_Pre-sampling</code> (Early Stop by Anima, Handover to Krea 2 for Remaining Denoising)</h3><p><strong>Mechanism</strong>: Anima stops sampling at an intermediate step of the denoising process and hands over the partially converged latents to Krea 2, which then completes all subsequent denoising steps.</p><p>- <strong>Advantage</strong>: The majority of denoising is executed by Krea 2, so the final image aligns more closely with Krea 2's generative style. Additionally, the character LoRA retains more flexibility in the intermediate latent space, leading to better reconstruction of clothing and other fine details.</p><p>- <strong>Disadvantage</strong>: Because Anima does not run to completion, Krea 2 still bears a significant denoising burden, and the risk of structural collapse or artifacts remains higher than in Variant 1, making it less stable overall.</p><hr /><h3>核心思想</h3><p>以 Anima 模型作为姿态控制模型（pose-conditioning model），生成带有姿态条件信息的初始噪声潜变量（conditioned latent noise），供 Krea 2 模型作为去噪过程的起始输入。</p><h3 id="krea2_with_anima_samplinganima-krea-2">变体一<code>Krea2_with_Anima_Sampling</code>（Anima 完整采样，Krea 2 精修）</h3><p><strong>机制</strong>：Anima 模型完成完整的去噪采样（denoising sampling）流程，输出最终的潜变量（latents）后，交由 Krea 2 模型进行后续精修。</p><p>- <strong>优势</strong>：因 Anima 全程主导去噪，构图与整体结构高度可控，生成结果稳定一致。</p><p>- <strong>劣势</strong>：构图在 Anima 阶段已完全定型，角色 LoRA 对服饰细节的约束力在后期精修中难以有效体现，导致服饰还原度较低。</p><h3 id="krea2_with_anima_pre-samplinganima-krea-2">变体二<code>Krea2_with_Anima_Pre-sampling</code>（Anima 中途停止，Krea 2 接管剩余去噪）</h3><p><strong>机制</strong>：Anima 模型在去噪过程的中间步骤提前终止采样，将未完全收敛的潜变量交接给 Krea 2 模型，由 Krea 2 完成后续的全部去噪步骤。</p><p>- <strong>优势</strong>：大部分去噪过程由 Krea 2 执行，最终图像更贴近 Krea 2 的生成风格；同时角色 LoRA 在中间潜空间拥有更多调整余地，服饰等细节还原更佳。</p><p>- <strong>劣势</strong>：由于 Anima 未完成全程，Krea 2 仍需承担较重的去噪任务，图像出现结构紊乱、伪影等崩坏现象的风险依然存在，稳定性不及变体一。</p><hr /><p></p><h3 id="krea2_with_anima_sampling">Krea2_with_Anima_Sampling<br /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/08fd216b-4507-4cfa-b756-4d6cf9580b29/original=true/08fd216b-4507-4cfa-b756-4d6cf9580b29.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/024c4d71-48b1-42bc-9c4c-53d0537e3046/original=true/024c4d71-48b1-42bc-9c4c-53d0537e3046.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/9d5d337a-5f8b-4c47-80da-84255de63b58/original=true/9d5d337a-5f8b-4c47-80da-84255de63b58.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d53622e2-e2a1-489b-92a9-593c6658963a/original=true/d53622e2-e2a1-489b-92a9-593c6658963a.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/22b29b9c-bb18-4753-ae56-f1dca4c9a313/original=true/22b29b9c-bb18-4753-ae56-f1dca4c9a313.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/b35219ce-2179-4809-860d-99c6d12d9e2d/original=true/b35219ce-2179-4809-860d-99c6d12d9e2d.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/83f62e92-4976-41f5-a601-0769bd1158e0/original=true/83f62e92-4976-41f5-a601-0769bd1158e0.jpeg" /></h3><h3 id="krea2_with_anima_pre-sampling">Krea2_with_Anima_Pre-sampling<br /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/26094a7e-dfe3-4e83-bcaf-3e8991297d3a/original=true/26094a7e-dfe3-4e83-bcaf-3e8991297d3a.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6b7a2e37-7d60-4ce1-8fe3-11aec014295f/original=true/6b7a2e37-7d60-4ce1-8fe3-11aec014295f.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/b6bbc0c6-5e83-4d89-8b52-89c823a14ec9/original=true/b6bbc0c6-5e83-4d89-8b52-89c823a14ec9.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/78705ccb-115e-46f9-ac05-bdc597ded341/original=true/78705ccb-115e-46f9-ac05-bdc597ded341.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ea8ffda4-a7f8-478d-985d-7ed0fee95a99/original=true/ea8ffda4-a7f8-478d-985d-7ed0fee95a99.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/cf14949e-a6c9-479a-85fe-dfb9d5c8e24f/original=true/cf14949e-a6c9-479a-85fe-dfb9d5c8e24f.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2750be85-e1bb-46cb-8ac4-4119767cd100/original=true/2750be85-e1bb-46cb-8ac4-4119767cd100.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/3e11a0d7-29cd-4763-b1a1-4f73f9649a02/original=true/3e11a0d7-29cd-4763-b1a1-4f73f9649a02.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d8e9b400-227d-401b-b247-bd751f73bae5/original=true/d8e9b400-227d-401b-b247-bd751f73bae5.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8a169d8c-b967-4b24-a56b-d64b56c4fb2a/original=true/8a169d8c-b967-4b24-a56b-d64b56c4fb2a.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/640b3f2b-6a37-4d83-98f1-a111faebc8fa/original=true/640b3f2b-6a37-4d83-98f1-a111faebc8fa.jpeg" /></h3><p></p>

Krea2_with_Anima_Sampling

<p><strong>Mechanism</strong>: Anima stops sampling at an intermediate step of the denoising process and hands over the partially converged latents to Krea 2, which then completes all subsequent denoising steps.</p><ul><li><p><strong>Advantage</strong>: The majority of denoising is executed by Krea 2, so the final image aligns more closely with Krea 2's generative style. Additionally, the character LoRA retains more flexibility in the intermediate latent space, leading to better reconstruction of clothing and other fine details.</p></li><li><p><strong>Disadvantage</strong>: Because Anima does not run to completion, Krea 2 still bears a significant denoising burden, and the risk of structural collapse or artifacts remains higher than in Variant 1, making it less stable overall.</p></li></ul>

Files with SHA256: 2d627b86ae6bc4df690b6f37db9dd5959b9b90e6d767dea4e169ec7253209c94