Full fine-tune of the MOSS-Image-Base DiT (text encoder and VAE frozen) with DiffSynth-Studio, 1 epoch over 11,264 caption rows, lr 1e-5, 1 MP aspect-bucketed.
checkpoints/step-N.safetensors: bf16 DiT, same keys as the base transformer/ (loads in ComfyUI as is)int8_convrot/step-N_int8_convrot.safetensors: INT8 ConvRot, same layout as Comfy-Org/MOSS-Image moss_image_base_int8_convrotUse with the Comfy-Org MOSS-Image Base workflow (Qwen3.5 text encoder, FLUX.2 VAE).
checkpoints/step-1000.safetensors 10 GBcheckpoints/step-2000.safetensors 10 GBcheckpoints/step-3000.safetensors 10 GBcheckpoints/step-4000.safetensors 10 GBcheckpoints/step-5000.safetensors 10 GBcheckpoints/step-5632.safetensors 10 GBint8_convrot/step-1000_int8_convrot.safetensors 5.4 GBint8_convrot/step-2000_int8_convrot.safetensors 5.4 GBint8_convrot/step-3000_int8_convrot.safetensors 5.4 GBint8_convrot/step-4000_int8_convrot.safetensors 5.4 GBint8_convrot/step-5000_int8_convrot.safetensors 5.4 GBint8_convrot/step-5632_int8_convrot.safetensors 5.4 GBOpenMOSS-Team/MOSS-Image-Base