Follow-up to my 4MP MOSS-Image post: Quick 3s local MiniMax H3 I2V tests.
retained text
I wanted to see how yesterday's raw 4MP MOSS-Image outputs behave in a local MiniMax H3 Image-to-Video pipeline before I had to head out. Kept the renders short (\~3s each) for a fast turnaround, expecting some morphing, but the detail retention and temporal consistency honestly surprised me. Pipeline & Specs: Base Frames: MOSS-Image (5B) Native 4MP raw outputs (from my previous post) Video Model: Local MiniMax H3 (Image-to-Video) Hardware: NVIDIA RTX PRO 6000 Prompts: Dead simple 3–4 word prompts focusing only on expression/attitude to see how the model reacts to the base image alone. What stood out: Even in a quick test, feeding native 4MP micro-textures into MiniMax H3 makes a massive difference. The suede jacket texture, skin pores, chain, tweed fabric, and thin glasses frames stay completely locked without any texture sliding. Most importantly, on the second clip, the model respected the stylized elongated neck anatomy instead of trying to force-correct it back into standard human proportions during motion. Base image workflow is the standard built-in ComfyUI MOSS template: https://docs.comfy.org/tutorials/image/moss-image/moss-image Edit: Uploaded the full 4MP photorealism showcase reel to YouTube Shorts—I'll also be posting the complete MOSS-Image to MiniMax H3 local tutorial on the channel soon: https://youtube.com/shorts/-hn_QRt7sig
document fetched 2026-10-26 00:38 · canonical https://reddit.com/r/comfyui/comments/1xev4vb/followup_to_my_4mp_mossimage_post_quick_3s_local