Diffusion models

OpenTernary can also quantize a component of a Diffusers pipeline, such as the transformer or UNet, and generate images under a fixed protocol for comparison. Install the extra first:

uv tool install "openternary[diffusion]" --python 3.12 --reinstall   # diffusers==0.40.0
openternary setup                                                    # the reinstall removed the GPU build

Quantize a component

A pipeline folder contains model_index.json and one subfolder per component. Choose the component explicitly:

openternary plan ./my-pipeline --component transformer
openternary quantize ./my-pipeline --component transformer --device cpu --output runs/pipe-transformer

Only recognizable two-dimensional linear weights are converted. Convolution weights and the VAE are left alone. This is not low-bit support for whole pipelines. Output shards use Diffusers' standard names, so Diffusers can load the component back.

Evaluate with a fixed profile

Image quality is judged by people, so evaluate makes generation reproducible and records evidence, but never passes an image as "good". Describe the protocol in a profile:

diffusion-profile.json

{
  "schema_version": 1,
  "family": "diffusion",
  "name": "fixed-prompt-v1",
  "prompts": ["A red cube on a white table"],
  "seeds": [42],
  "steps": 20,
  "width": 512,
  "height": 512,
  "guidance_scale": 3.5,
  "warmup": true
}

openternary evaluate ./my-pipeline --profile diffusion-profile.json --dry-run      # check the profile
openternary evaluate ./my-pipeline --profile diffusion-profile.json --output runs/pipe-original

To evaluate the pipeline with your quantized component swapped in, add "component": "transformer" to the profile and pass the artifact:

openternary evaluate ./my-pipeline --profile diffusion-profile.json \
    --component-artifact runs/pipe-transformer/artifacts/snapshot --output runs/pipe-ternary

The replacement must currently be safetensors.

What is recorded

evaluation.json stores the prompts, seeds, steps, resolution and the pipeline's scheduler configuration, and, for each image, its hash, the generation time and pixel statistics. Comparisons check that these settings match.

Every result carries human_review_required: pixel statistics and hashes can show that images changed, but not whether they got worse. OpenTernary therefore never marks a diffusion result as quality-accepted.

What has been verified

A synthetic Stable Diffusion pipeline at 32×32 pixels and 2 steps has been converted with its UNet swapped in, and has generated finite images, on Windows and WSL CPUs and on a Radeon RX 9070 XT in FP32 and BF16. No real pretrained pipeline and no image quality have been evaluated. See Status.