Entradas
Nota: With
"ai-toolkit" and "ai-toolkit (kv_cache off)", or when no VAE is connected and image_strength is above 0, the text encoder also sees every reference, downscaled to at most 384x384 pixels and preceded by a Picture N: marker in the prompt, so the instruction can refer to them by number. With "Moss-Image-Edit" and a VAE connected, the references reach the model only as reference latents and the text encoder sees the instruction alone. "ai-toolkit (kv_cache off)" additionally keeps the references attending to the image being generated, so they are recomputed on every sampling step.
Salidas
Esta documentación fue generada por IA. Si encuentra algún error o tiene sugerencias de mejora, ¡no dude en contribuir! Editar en GitHub
Source fingerprint (SHA-256):
673f098df6a8e466a5c8060009f446eac8f7556311a870de626c5f4bfc3ab196