Krea2 is still unbeatable so far.Convert anime image to realistic using moss as edit model, thanks to ostris.Suggest using with <a target="_blank" rel="ugc" href="https://civarchive.com/models/3012032/moss-edit-anime2real-workflow-anime-to-realistic">https://civarchive.com/models/3012032/moss-edit-anime2real-workflow-anime-to-realistic</a>Workflow embedded in example images.ai-toolkit training config file in the Optional Files.Dependencies:custom node: <a target="_blank" rel="ugc" href="https://github.com/ostris/ComfyUI-MOSS-Image-Ostris-Edit">https://github.com/ostris/ComfyUI-MOSS-Image-Ostris-Edit</a>Since ComfyUI has refactored the MossImage code, please use my modified version instead: <a target="_blank" rel="ugc" href="https://github.com/SurpassHR/ComfyUI-MOSS-Image-Ostris-Edit">https://github.com/SurpassHR/ComfyUI-MOSS-Image-Ostris-Edit</a>Alternatively, you can make modifications yourself:Edit <code>ComfyUI/custom_nodes/ComfyUI-MOSS-Image-Ostris-Edit/nodes.py</code>Locate lines 161–162:Replace the function<pre><code>def _pe(dit, ids):
 return dit.pe_embedder(ids.unsqueeze(0)).transpose(1, 2).contiguous()</code></pre>with<pre><code>def _pe(dit, ids):
 if hasattr(dit, "pe_embedder"):
 return dit.pe_embedder(ids.unsqueeze(0)).transpose(1, 2).contiguous()
 ids_b = ids.unsqueeze(0)
 pe = torch.cat([comfy.ldm.flux.math.rope(ids_b[..., i], d, 10000) for i, d in enumerate(dit.rope_axes_dim)], dim=-3)
 return pe.unsqueeze(1).transpose(1, 2).contiguous()</code></pre>

[Moss Edit] Anime2Real (Anime to Realistic)

Trained 3,000 steps on 80 pairs of images.

Files with SHA256: 0af9a7c6d72431778ee3afbbf047ddcb1a9718875a599d443c3e314ff2a47614