The two images below were generated using the same stage c conditioning latents, and the same sampler/scheduler (euler, simple, 10 steps). The first image was generated using ComfyUI, the 2nd image using the native diffusers stage b implementation. This behaviour appears consistently across different samplers, schedulers and with different stage c conditioning. Using the diffusers implementation of stage a to decode ComfyUI stage b output latents does produce artifacts, suggesting that the issue is specifically in the stage b implementation, not stage a.
The two images below were generated using the same stage c conditioning latents, and the same sampler/scheduler (euler, simple, 10 steps). The first image was generated using ComfyUI, the 2nd image using the native diffusers stage b implementation. This behaviour appears consistently across different samplers, schedulers and with different stage c conditioning. Using the diffusers implementation of stage a to decode ComfyUI stage b output latents does produce artifacts, suggesting that the issue is specifically in the stage b implementation, not stage a.
Comfy workflow and example latents (had to zip, github doesn't like the format) workflow_test.json test_latent.zip