Maximizing Identity Preservation in MiniMax H3 img2vid
NetworkSpecial3268 · reddit · 2026-08-10
Exploring Identity Preservation in H3 img2vid
In the ComfyUI community, users are discussing how to maximize identity preservation when using the MiniMax H3 image-to-video workflow.
Current Consensus & Bottleneck
- Ref2vid Workflow: The current consensus for the best identity preservation is using the Ref2vid workflow with a close-up character sheet and setting refimagesize to max.
- Performance Hit: However, this causes a significant slowdown in generation speed.
Alternative Approaches Explored
Users are exploring tricks to improve consistency in pure img2vid mode, such as: adding a "picture-in-picture" close-up of the face into the main starting image with prompting, or starting at a very high resolution like 2048x2048 but targeting a much smaller video resolution.
More from Multimodal
- DAPI Launches: A Video Editing CLI Built Exclusively for AI Coding Agents — Vjeux · 2026-08-11
- H3 Audio Generation Workflow: Achieving 48kHz High-Fidelity Upsampling via AudioSR — AppointmentOk972 · 2026-08-11
- ComfyUI Update: Improved H3 VAE Routines — PumpkinLeather8421 · 2026-08-11
- NVIDIA Releases Open-Weight Magpie TTS for Low-Latency Multilingual Voice Agents — Hugging Face Blog · 2026-08-11
- MiniMax H3 Team AMA Summary: 2K Regeneration Model and Sparse Attention Coming Soon — the_bollo · 2026-08-11
- Cinematic AI Reconstruction: The Night Abraham Lincoln Was Assassinated — Hamlet517 · 2026-08-11