import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("AcademiaSD/MiniMax-H3-Image-NF4-for-LoRA-Training", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]MiniMax-H3-Image NF4 for LoRA Training
Pre-quantized MiniMax-H3-Image packaged for AcademiaSD MiniMax-H3 Image LoRAlab, part of AcademiaSD LoRAlab Trainer Studio, a LoRA trainer for consumer GPUs. It is downloaded automatically by the LoRAlab; you do not need to fetch it by hand.
This is not an inference checkpoint. The transformer is stored as a per-layer NF4 cache that the LoRAlab rebuilds directly; it does not load with
from_pretrained. For generation, use the original model or its ComfyUI release.
What is inside
| Folder | Size | Content |
|---|---|---|
transformers/ |
17.23 GB | 33B single-stream DiT as an NF4 cache, one file per layer (66.3 GB in BF16) |
vae/ |
2.34 GB | Original image VAE (CNN encoder, 2D ViT decoder, 24 channels), diffusers format, unchanged |
scheduler/, pdd_config.json, modular_model_index.json |
— | Original files, unchanged |
LICENSE, NOTICE |
— | MiniMax-H3-Image Community License Agreement and the notice it requires |
Two folders come from AcademiaSD/MiniMax-H3-NF4, not from here, and the LoRAlab downloads only those two (35.6 GB of the 41.4 GB repo; nothing if you already train MiniMax-H3 video LoRAs):
text_encoder/(18.4 GB): MiniMax-H3-Image conditions on the unchanged Qwen3-VL text encoder of MiniMax-H3. Used by the pre-cache only.transformers/(17.2 GB): the MiniMax-H3 DiT the LoRA is trained on (see below).
Transformer (NF4)
- Every linear of the 50 blocks and of the 2 token-refiner blocks is NF4 (bitsandbytes, double quantization, BF16 compute): 373 layers.
proj_in, the timestep embedder,norm_outand the six per-step output heads (pdd_proj_out.0–5) are also stored unquantized (pc-*files): the model declares them precision-critical, and the heads are what make the 6-step distillation work.- Norms and other non-linear parameters are kept as in the original (
other-*files). - Measured weight error of the NF4 layers: 7–9 % relative to BF16. The precision-critical copies are bit-identical to the original.
- Generation from this cache, on an RTX 5080 with 16 GB: 1024×1024 in 6 steps in ~33 s, with a VRAM peak of 1.8 GB.
VAE
vae/ is the official vae_diffusers/ folder, byte-identical (verified by SHA-256). The LoRAlab
encodes every training image with it, never with the MiniMax-H3 video VAE, using the same pixel
normalization as the released pipeline.
What the LoRAlab trains with it
- Characters, objects and styles: one image + caption per sample.
- Trained on MiniMax-H3, made for MiniMax-H3-Image. MiniMax-H3-Image is distilled to 6 sampling steps with one output head per step and no CFG, and fine-tuning it directly fails: in our tests plain flow matching turned its high-noise heads into the dataset's average within ~100 steps (blurred, blotchy images, even at lr 1e-4), and anchoring those heads kept the images sharp but learned no likeness. MiniMax-H3-Image was fine-tuned from MiniMax-H3, which is not distilled and shares its latent space and every module a LoRA touches. So the LoRAlab trains the LoRA on the MiniMax-H3 DiT, with latents from this repository's VAE, and renders the training previews with MiniMax-H3-Image (6 steps, CFG 1) with the LoRA attached. Measured: previews stay sharp and pick up the dataset from the first few hundred steps.
- Low VRAM: transformer blocks are streamed to the GPU on demand. Measured on a 16 GB card: 30 of the 50 blocks resident, 13.9 GB peak with previews included, ~4.5 s per step at ~1 MP.
Exported LoRAs use the key layout of the original checkpoints (fused qkv_proj, interleaved per
attention head; mlp.fc1 with its SwiGLU halves in the original order), verified row by row
against the official MiniMax-H3-Image transformer/ and MiniMax-H3 FL2VA/transformer. The same
file applies to both models.
License
MiniMax-H3-Image is licensed under the MiniMax-H3-Image Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.
This repository is a modified redistribution (NF4 quantization) distributed under that same
agreement; see LICENSE and NOTICE. Among its terms:
- Use is only authorized within the Applicable Territory, which excludes the European Union, the United Kingdom, the Republic of Korea and the United States.
- Commercial use above USD 1M of annual revenue, or as a hosted service for third parties, requires a separate authorization from MiniMax.
- Use is subject to the Acceptable Use Policy (Exhibit A of the license).
By downloading these files you accept the terms of that license. This summary is not legal advice;
the LICENSE file is what applies.
Powered by MiniMax-H3-Image.
Credits
- MiniMax for MiniMax-H3-Image and MiniMax-H3.
- NF4 conversion and LoRAlab by AcademiaSD: YouTube · X · Discord · Ko-Fi
Resumen en español
MiniMax-H3-Image cuantizado a NF4 para el LoRAlab de AcademiaSD, que lo descarga solo.
Transformer de 33B en NF4 (17,2 GB en lugar de 66,3 GB) con las seis cabezas de salida por paso
sin cuantizar, y el VAE de imagen original. El text encoder no va aquí: es el Qwen3-VL de
MiniMax-H3. De AcademiaSD/MiniMax-H3-NF4 el LoRAlab descarga solo text_encoder/ y
transformers/: el LoRA se entrena sobre el DiT de MiniMax-H3 (sin destilar) con latentes del VAE
de este repo, y las previews se hacen con MiniMax-H3-Image, porque entrenar directamente el modelo
destilado lo emborrona o no aprende. Sirve para LoRAs de personajes, objetos y estilos. No es un
checkpoint de inferencia. Licencia MiniMax-H3-Image Community License:
uso no autorizado en la UE, Reino Unido, Corea del Sur ni EE. UU.
- Downloads last month
- 31