Instructions to use tsolful/Minimax_H3_Image_INT6_INT4GPTQ with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusion Single File
How to use tsolful/Minimax_H3_Image_INT6_INT4GPTQ with Diffusion Single File:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
MiniMax H3 Image (INT6 W6A8 & INT4 W4A8 GPTQ) for ComfyUI
Low-bit GPTQ quantizations of the MiniMax H3 Image diffusion transformer, packaged as single files for ComfyUI.
- Original model: MiniMaxAI/MiniMax-H3-Image
- Official ComfyUI repackage (pruned INT8 ConvRot model, VAE): Comfy-Org/MiniMax-H3-Image
Both files were quantized from the pruned BF16 weights (minimax_h3_image_pruned_bf16.safetensors) used by the ComfyUI repackage.
Files
| File | Format | Size | Notes |
|---|---|---|---|
minimaxh3_image_W6A8_GPTQ.safetensors |
INT6 W6A8, GPTQ | 15.61 GB | Highest quality |
minimaxh3_image_W4A8_GPTQ.safetensors |
INT4 W4A8 + W6A8 tail, GPTQ | 12.77 GB | Smallest. Last 20 blocks kept at W6A8 |
*_quant_report.json |
β | β | Per-layer quantization reports, for reference only |
W6A8 GPTQ: all transformer blocks are stored at 6-bit weights (group size 32) with 8-bit activations.
W4A8 GPTQ (mixed): blocks 0β29 are stored at 4-bit weights (asymmetric codebook, group size 16). The last 20 blocks (30β49) are kept at W6A8 (group size 16). That adds size over a pure 4-bit file but keeps the late blocks, where 4-bit error is largest, close to the W6A8 file.
Text encoder and VAE
These are not included here. The text encoder is shared with MiniMax H3:
- Text encoder: Comfy-Org/MiniMax-H3 βΊ text_encoders, for example
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors - VAE:
vae/minimax_h3_image_vae_fp16.safetensorsfrom Comfy-Org/MiniMax-H3-Image
Model storage paths
π ComfyUI/
βββ π models/
β βββ π diffusion_models/
β β βββ minimaxh3_image_W6A8_GPTQ.safetensors
β β βββ minimaxh3_image_W4A8_GPTQ.safetensors
β βββ π text_encoders/
β β βββ qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
β βββ π vae/
β βββ minimax_h3_image_vae_fp16.safetensors
You only need the diffusion model file you plan to use.
Load the diffusion model with ComfyUI's Load Diffusion Model node, the same way as the Comfy-Org file. The quantization settings are stored in ComfyUI's native _quantization_metadata header, so no extra loader is needed. Requires comfy-kitchen >= 0.2.37 on ComfyUI v0.38.0 or newer.
Conversion details
The transformer has 50 blocks (0β49). Each has four large linear layers: attn.qkv_proj, attn.out_proj, mlp.fc1 and mlp.fc2. These are the layers being quantized. Small, sensitive parts such as embedders, modulation, output layer and norms stay at full precision. Converted with my ComfyUI-INT8-Fast (fork).
| Setting | W6A8 GPTQ | W4A8 GPTQ |
|---|---|---|
| Weight format | Symmetric INT6 | Asymmetric INT4 codebook, W6A8 on blocks 30β49 |
| Group size | 32 | 16 (both the INT4 and W6A8 parts) |
| Scale dtype | FP8 E4M3 | FP8 E4M3 |
| GPTQ-calibrated layers | 192 | 200 (includes the token refiner) |
| Token refiner | Not quantized | INT8 |
| GPTQ settings | 2 passes, damp 0.02, act-order, sequential | Same |
| Calibration tokens per layer | 75,648 | 75,648 |
Layer-level improvement from GPTQ
The mean output error per layer is measured against BF16 on the calibration data. GPTQ is compared with plain rounding (RTN) at the same bit width.
| Format | Plain rounding (RTN) | GPTQ | Reduction |
|---|---|---|---|
| W6A8 | 0.0190 | 0.0060 | 68 % |
| W4A8 (mixed) | 0.0385 | 0.0113 | 71 % |
GPTQ cuts per-layer error to roughly a third of plain rounding for both files.
End-to-end error per denoising step
This is the relative RMSE of the full denoiser output vs BF16 at each sampling step (lower is better).
| Step | Sigma | W6A8 GPTQ | W4A8 GPTQ |
|---|---|---|---|
| 0 | 1.000 | 0.035 | 0.040 |
| 1 | 0.968 | 0.086 | 0.070 |
| 2 | 0.888 | 0.043 | 0.051 |
| 3 | 0.750 | 0.029 | 0.031 |
| 4 | 0.534 | 0.011 | 0.013 |
| 5 | 0.233 | 0.004 | 0.004 |
| Mean | 0.035 | 0.035 |
Cosine similarity with BF16 stays at 0.996 or higher at every step for both files.
- Both files track BF16 closely. The error is largest in the first, noisiest steps and drops below 1.3 % by step 4.
- The mixed W4A8 file is almost level with W6A8 end to end while being 2.8 GB smaller. Keeping the last 20 blocks at W6A8 recovers most of the 4-bit loss.
- Measured on calibration data. These numbers come from a single run on the data used for calibration, with no held-out set, so treat them as indicative. Small differences between the files, such as step 1, are within that noise.
Check output quality for your own prompts and workflows before relying on it.
License
These files are derived from MiniMax H3 Image and are released under the MiniMax H3 Community License Agreement. See the original model repository for the full terms. Use of these files is subject to that agreement.
- Downloads last month
- 435
Model tree for tsolful/Minimax_H3_Image_INT6_INT4GPTQ
Base model
MiniMaxAI/MiniMax-H3