Original character LoRA for MiniMax H3 learnings. Workflows included.
So I attempted to create a LoRA of an original character, using 8 Z-Image Turbo stills, a custom ElevenLabs voice, and 14 short clips of her talking (from 5 Seedance gens with the Z-Image refs).
Proof1 community thread
Key points
- H3 t2v is just so damn good looking compared to ref2va/fl2va and even RefMod.
- Left is H3 with a RefMod (photos and voice sample).
- Rendering on an RTX 3090: 864x480 at 8 steps with the DMAD 4-step LoRA, 3D latent upscale to 2 MP, 2-step refine at 0.4 denoise, then DLSS 5 at 0.8 intensity.
- Workflows (API format, drag onto the canvas): H3 + LoRA, H3 + LoRA + RefMods on ref2va, DLSS 5, and both RefMod builders:
Sources (1)
- [1]Original character LoRA for MiniMax H3 learnings. Workflows included.r/StableDiffusion (top, day) · Oct 21, 02:14 PM
So I attempted to create a LoRA of an original character, using 8 Z-Image Turbo stills, a custom ElevenLabs voice, and 14 short clips of her talking (from 5 Seedance gens with the Z-Image refs).
H3 t2v is just so damn good looking compared to ref2va/fl2va and even RefMod.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 20, 2026Gradium Open-Sources Phonon, a 31M-Parameter On-Device TTS Model Under Apache 2.0
- Oct 20, 2026How to train MOSS-Image LoRAs locally (8 GB VRAM): settings, tips and workflow
- Oct 20, 2026Voxtral 3
- Oct 20, 2026[AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback
- Oct 20, 2026Moss LoRAs Training
- Oct 19, 2026Amazon SageMaker JumpStart: 2026 year-to-date in review
