Same Krea 2 + MiniMax H3 realism pipeline reposted, asking what breaks the AI illusion
felandocruzel2000 · reddit · 2026-11-11
A Redditor reposted their realistic phone-footage AI video pipeline seeking critique: Krea 2 stills hold up, but H3 video softens eyes and shifts faces.
Image stage: ComfyUI + krea2turbobf16 (9 steps, CFG 1.0). Video stage: MiniMax H3 Ultra on RunPod with Qwen3-VL 32B text encoder, SelfLiftH3Sampler at 12 steps starting low-res then upscaling, DMAD v4 LoRA at 0.75; 15s renders in 5 minutes on an RTX PRO 4500. The prompt uses a 0–15s timeline (wide shot → walk closer → slow head turn) with fixed camera and deep depth of field.
Related event: Reddit Seeks Help Fixing Realism in Krea + MiniMax H3 Video Workflow(2 posts)→
More from Multimodal
- MiniMax H3 image model: 16MP in 15.7s, runs on 12GB GPUs — CeFurkan · 2026-11-11
- ACE Music v1 debuts at #1 on AA-Music-Vocal, ahead of Suno v6 — ArtificialAnlys · 2026-11-11
- Claude 3 Sonnet wrote three 'visions' before retiring, rendered verbatim in Midjourney — repligate · 2026-11-11
- fal Agent lands in Photoshop, bringing multi-model image and video generation in-app — gorkem · 2026-11-11
- Animated NYC Millennium Eve short film made with Google Flow's Veo and Omni — MoHassan_Film · 2026-11-11
- Claude writes all code from one photo: photorealistic sunlit room rendered in vork — ravisparikh · 2026-11-11