Harsha001's picture
Model card, licence, notice and stable-diffusion.cpp patch
dc51a9a verified
|
Raw History Blame Contribute Delete
8.84 kB
---
license: other
license_name: minimax-h3-image-community-license-agreement
license_link: LICENSE
base_model: MiniMaxAI/MiniMax-H3-Image
base_model_relation: quantized
pipeline_tag: text-to-image
library_name: gguf
tags: [gguf, rocmfpx, rocmfp4, amd, rocm, vulkan, strix-halo, stable-diffusion.cpp, text-to-image, minimax-h3-image, quantized]
extra_gated_heading: Confirm your location and accept the MiniMax-H3-Image Community License
extra_gated_prompt: >-
The MiniMax-H3-Image Community License Agreement grants rights only outside the Excluded Territories: the European
Union, the United Kingdom, the Republic of Korea and the United States of America. Use, reproduction, distribution
and outputs inside those territories are not licensed.
extra_gated_fields:
I confirm that I am not located in, and will not use these files or their outputs in, the European Union, the United Kingdom, the Republic of Korea or the United States of America: checkbox
I have read and accept the MiniMax-H3-Image Community License Agreement (LICENSE) and its Acceptable Use Policy: checkbox
extra_gated_button_content: Agree and access
---
# MiniMax-H3-Image, ROCmFPX GGUF
**Powered by MiniMax-H3-Image.**
Quantized files of [MiniMax-H3-Image](https://huggingface.co/MiniMaxAI/MiniMax-H3-Image), MiniMax's text-to-image and
image-editing model (6-step distilled, no CFG), for [stable-diffusion.cpp](https://github.com/leejet/stable-diffusion.cpp).
Each file is the diffusion transformer only. The text encoder and the VAE are the usual downloads (see [Run](#run)).
> **Licence and territory.** MiniMax-H3-Image is distributed under the MiniMax-H3-Image Community License Agreement
> (`LICENSE`, `NOTICE`). It grants rights **only outside the European Union, the United Kingdom, the Republic of Korea
> and the United States of America**. Do not download, use or share these files, or anything generated with them, in
> those territories. Use must also follow the Acceptable Use Policy in `LICENSE`. Annualized revenue above 1 million USD
> or offering the model as a hosted service needs MiniMax's prior written permission.
>
> Not made, sponsored or endorsed by MiniMax.
> **These files have not been tested on AMD hardware yet.** Please report speed and quality in the Community tab.
>
> The `ROCMFPX` and `ROCMFP4` files need stable-diffusion.cpp built on the ROCmFPX fork of ggml with the small patch
> in this repo (`sdcpp-rocmfpx.patch`). The `Q6_K` and `Q5_K` files are stock llama.cpp weight types and load in
> stable-diffusion.cpp from commit `32d516c` (MiniMax-H3-Image support) on.
## Files
| File | Size | LPIPS | PSNR | SSIM | |
|---|---:|---:|---:|---:|---|
| `MiniMax-H3-Image-Q8_0_ROCMFPX.gguf` | 20.8 GB | **0.032** | 31.05 | 0.957 | 8-bit, closest to BF16 |
| `MiniMax-H3-Image-Q6_K.gguf` | 16.6 GB | 0.060 | 25.48 | 0.916 | 6-bit, stock type, our importance matrix |
| `MiniMax-H3-Image-Q5_K.gguf` | 13.9 GB | 0.113 | 22.30 | 0.847 | 5-bit, stock type, our importance matrix |
| `MiniMax-H3-Image-Q4_0_ROCMFP4.gguf` | 11.4 GB | 0.193 | 19.52 | 0.757 | 4-bit, smallest |
Compared with GGUFs already on the Hub and with stock files, measured the same way:
| Not published | Size | LPIPS | PSNR | SSIM |
|---|---:|---:|---:|---:|
| [Hearthscale](https://huggingface.co/Hearthscale/MiniMax-H3-Image-GGUF) Q8_0 | 21.4 GB | 0.041 | 30.34 | 0.945 |
| Stock Q8_0 (stable-diffusion.cpp convert) | 21.4 GB | 0.043 | 30.07 | 0.940 |
| ROCmFPX 6-bit | 16.4 GB | 0.085 | 24.43 | 0.883 |
| Stock Q6_K without importance matrix | 16.6 GB | 0.080 | 24.25 | 0.891 |
| [Hearthscale](https://huggingface.co/Hearthscale/MiniMax-H3-Image-GGUF) Q5_K_M | 13.9 GB | 0.124 | 21.65 | 0.834 |
| [Hearthscale](https://huggingface.co/Hearthscale/MiniMax-H3-Image-GGUF) Q4_K_M | 11.4 GB | 0.202 | 18.54 | 0.740 |
| Stock Q4_K, our importance matrix | 11.4 GB | 0.218 | 18.46 | 0.730 |
At every size the files here are closer to BF16 than the Hub GGUFs of the same size. The ROCmFPX 8-bit file is also
0.6 GB smaller than a Q8_0. At 6-bit the stock `Q6_K` with our importance matrix beat the ROCmFPX 6-bit type, so the
`Q6_K` is the one here. Abiray's GGUFs could not be compared: they store the output head in a ComfyUI-specific layout
that stable-diffusion.cpp does not load.
**How this was measured:** 12 prompts (portrait, text on signs, watercolor, still life, animals, aerial, night market and
others) at 512x512, seed 1, 6 Euler steps with the MiniMax-H3-Image sigma schedule, rendered with stable-diffusion.cpp on
an NVIDIA GB10 GPU after expanding each file to F16, and compared image by image with the same render from the BF16
model. LPIPS is how different two images look to a neural network (lower is better); PSNR and SSIM compare the pixels
(higher is better). A 6-step model turns small weight differences into small changes of composition, so these numbers
measure how far each file drifts from BF16 rather than how good its images look.
**Running the real files on a CPU.** The scores above use F16 expansions of the files, so they measure the weights. To
check the runtime path too, the real 8-bit and 4-bit files rendered 2 of the prompts on the test machine's ARM CPU:
| File | LPIPS, CPU, real file | LPIPS, same 2 prompts, F16 expansion |
|---|---:|---:|
| `Q8_0_ROCMFPX` | 0.097 | 0.054 |
| Stock Q8_0 | 0.050 | 0.049 |
| `Q4_0_ROCMFP4` | 0.199 | 0.185 |
| Stock Q4_K | 0.188 | 0.184 |
On the CPU, the ROCmFPX files drifted further from BF16 than their F16 expansions did, while the stock files did not.
The cause has not been found yet. The CPU dot products themselves are exact integer sums, like stock `Q8_0`. So if you
run on a CPU, the `Q6_K` and `Q5_K` files here, or a stock Q8_0, are the safer choice for now. The ROCmFPX files are
meant for AMD GPUs, where they run on different kernels that have not been tested with this model yet. Two prompts is
a small sample.
## How these were made
1. Source: Comfy-Org's `minimax_h3_image_pruned_bf16.safetensors` (revision `892de7c`), the file stable-diffusion.cpp's
docs use. In it, MiniMax's timestep adaLN network is replaced by a small precomputed table and the six PDD output
heads are fused, which is why it is 40 GB instead of the original 66 GB. The Hub GGUFs above are made from the same file.
2. **Importance matrix.** 20 renders of 16 calibration prompts (none of them the test prompts) with the BF16 model,
recording the inputs of every linear layer. stable-diffusion.cpp writes the matrix with a `model.diffusion_model.`
prefix that its own converter does not look up for a stand-alone diffusion file, so without renaming the entries
it is silently ignored. The entries were renamed so that the matrix is actually used.
3. **Tensor scales** (ROCmFPX files only). ROCmFPX block scales have a lower limit, and many MiniMax-H3-Image weights
are small enough to hit it. Each attention and feed-forward matrix is multiplied by a power of two before it is
quantized, and the inverse is stored as `<name>.weight_scale`, which stable-diffusion.cpp already applies after the
matrix multiplication.
4. `sd-cli -M convert --type <type> --imat-in <matrix>`. The 208 attention and feed-forward matrices take the named type.
Norms and the small input/output projections stay in F32, the text projection in Q8_0, and the rest as in the source.
## Run
```bash
git clone https://github.com/HarshaVardhan3002/ROCmFPX
git clone https://github.com/leejet/stable-diffusion.cpp && cd stable-diffusion.cpp
git checkout 32d516c && git submodule update --init --recursive
git apply ../sdcpp-rocmfpx.patch
cmake -B build -DCMAKE_BUILD_TYPE=Release -DSD_USE_UPSTREAM_GGML=ON -DSD_GGML_SOURCE_DIR=$PWD/../ROCmFPX/ggml -DGGML_VULKAN=ON
cmake --build build -j --target sd-cli
```
Text encoder [`qwen3vl_32b_minimax_h3-Q4_K_M.gguf`](https://huggingface.co/leejet/MiniMax-H3-GGUF) and VAE
[`minimax_h3_image_vae_fp16.safetensors`](https://huggingface.co/Comfy-Org/MiniMax-H3-Image/tree/main/vae):
```bash
build/bin/sd-cli -M img_gen --diffusion-model MiniMax-H3-Image-Q8_0_ROCMFPX.gguf \
--llm qwen3vl_32b_minimax_h3-Q4_K_M.gguf --vae minimax_h3_image_vae_fp16.safetensors \
--cfg-scale 1.0 --sampling-method euler --steps 6 --sigmas minimax-h3-image -W 1024 -H 1024 -o out.png \
-p "A shop sign that reads 'OPEN LATE' above a small blue door, rainy street, reflections on wet cobblestones."
```
See stable-diffusion.cpp's MiniMax-H3-Image guide for image editing.
## Credits and license
- Model: [MiniMaxAI/MiniMax-H3-Image](https://huggingface.co/MiniMaxAI/MiniMax-H3-Image), MiniMax-H3-Image Community License Agreement.
- Pruned BF16 file: [Comfy-Org/MiniMax-H3-Image](https://huggingface.co/Comfy-Org/MiniMax-H3-Image).
- Runtime: [stable-diffusion.cpp](https://github.com/leejet/stable-diffusion.cpp) by leejet.
See `LICENSE` and `NOTICE`.