AiAngelH3 is a weight merge of two MiniMax H3 checkpoints, tuned for realistic POV clips that keep a reference face:
40% H3 Eros Max beta5 TURBO-hybrid int8 by TenStrip
60% RedCraft H3 A2A-RED beta1 by AiMetatron
Both sources are int8 tensorwise + convrot, so the merge was done offline per tensor; the file is a drop-in diffusion model for ComfyUI's Load Diffusion Model node.
Recommended settings
Sampler euler, scheduler simple, 8 steps
No turbo LoRA (the Eros half carries its turbo-hybrid delta; stacking a turbo LoRA made results worse in our tests)
Concept LoRAs load well at low strength, e.g. Mystic XXX 0.4, HMCumshot 0.9
Reference faces: two clean real frames of the face as
<Picture 1>/<Picture 2>, scene described in text, gives the strongest likenessNeeds the standard H3 text encoder (
qwen3vl_32b_minimax_h3_nvfp4_awq), video VAE fp16 and audio VAE
Measured on an RTX PRO 6000: about 95 s for a 5 s 576×1024 clip, about 295 s for 15 s.
Run it on RunPod
No 96 GB GPU at home? The free AI Angel ComfyPod template (you pay only for the GPU and disk you rent on RunPod) boots ComfyUI with H3 ready and downloads AiAngelH3 at boot by default, so there is nothing to set. It also comes with a Dashboard where you fill in a form to make a clip, no nodes needed. Step-by-step guide: https://www.thepexcel.com/aiangel-comfypod/
New to RunPod: https://runpod.io?ref=fbrjus7u
Credits and license
Built from H3 Eros Max (tenstrip) and RedCraft (AiMetatron) — thank you both.
H3 Eros Max allows only selling generated images and running on Civitai's own generator, so the same applies here: do not sell this model or merges of it, and do not run it on paid generation services outside Civitai. Running it yourself, locally or on your own rented GPU, is fine.
NOTICE: This file has been modified from MiniMax H3 (a merge of two MiniMax H3 model derivatives). MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved. By downloading or using this model you agree to that license and its Acceptable Use Policy, which apply to you in full: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE The license does not cover the European Union, the United Kingdom, the Republic of Korea or the United States of America: do not download or use this model there. Anything you post publicly from it must be clearly labelled as AI-generated.
Adults only. Do not use it to depict real people without their consent, or anyone under 18.
Description
First release: 40% H3 Eros Max beta5 TURBO-hybrid int8 + 60% RedCraft H3 A2A-RED beta1, int8 tensorwise + convrot.
Recommended: euler / simple, 8 steps, no turbo LoRA.
FAQ
Comments (11)
Does is adresses the smearing issue and the distorted faces at distance? Thanks
getting good turbo results in four steps with this checkpoint with a 50/50 mix of tao3 step lora and fastvideo datafree preview 1 and er_sde simple. It doesn't have that over-saturated turbo look.
thanks for your feedback ❤️
If I understand correctly then "50/50 mix" means taomate lora 0.5 strength + fastvideo datafree preview 1 lora 0.5 strength?
I did 1.0 and 1.0 then 0.5 strength simple merge but also doing LatentMultiply 0.9 before vae decode can reduce contrast too, may be better.
@ShinyLilly Still don't really understand your setup.
do this model uses shift? it's not clear in the description
12/8 works fine for me.
I really like it, great work.
Thank you very much ❤️
Hi there, this model is excellent. I have long used Eros Beta5. Your model greatly addresses the shortcomings of Eros for Asian characters. Thank you for your contributions to the open-source community.