Original character LoRA for MiniMax H3 learnings. Workflows included.
retained text
H3 t2v is just so damn good looking compared to ref2va/fl2va and even RefMod. 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). Left is H3 with a RefMod (photos and voice sample). Right is the same seed with the LoRA at 0.8, followed by DLSS 5. The RefMod keeps her face consistent, but everything looks a bit too polished. With the LoRA, she looks closer to t2v quality. I trained her in Fizgig, running headless: rank 16, Automagic v3, Circlestone adapter, EMA 0.98, 40 epochs (3,520 steps) on the pruned int8 fl2va model. Dataset was the images plus face crops, 14 clips cut to the 17n+5 frame grid, and 45 voice clips trimmed to exact grid lengths. Training took 41 minutes on a rented H200 (free Modal credits). 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. About 4.5 minutes for an 8-second clip, plus 15 seconds for DLSS on a second card (5060 Ti). A few things I figured out: * LoRA at 0.8 is the sweet spot. At 1.0 she looked she was wearing too much foundation. Prompting "no makeup" somehow made it worse. * Don't trust the trainer previews. Late epochs looked terrible there, then great in the actual workflow. * RefMods pick up lighting from the reference photos. Studio lit stills made everything look like a moisturizer commercial, switching to candid photos fixed it. The strength slider did basically nothing. * The second pass loses the cloned voice. Save the stage-1 audio and put it back afterward. * Strong prior training data can still take over. Put her in an X-Files basement with Mulder and Scully, and suddenly she looks like Agent Reyes (the weird seasons) but with a British voice. Workflows (API format, drag onto the canvas): H3 + LoRA, H3 + LoRA + RefMods on ref2va, DLSS 5, and both RefMod builders: [https://drive.google.com/drive/folders/1T-ELnDiOQ2pOJOVaaL\_MvyfXcmeUdXeK?usp=sharing](https://drive.google.com/drive/folders/1T-ELnDiOQ2pOJOVaaL_MvyfXcmeUdXeK?usp=sharing) Props to shootthesound (Fizgig), Luisacaotica (RefMod nodes), Blueforcer and Merserk (DLSS 5 Enhancer), ZhengmingYu and Kijai (DMAD), and circlestone-labs (training adapter).
document fetched 2026-10-22 00:34 · canonical https://reddit.com/r/StableDiffusion/comments/1xbobrr/original_character_lora_for_minimax_h3_learnings