# Canter releases

[Model card](README.md) · [Getting started](README.md#getting-started) ·
[API and inference parameters](API.md)

Checkpoint tags are immutable snapshots. The installed package supplies the
inference code, so current code can load any compatible checkpoint tag without
executing the Python files stored in that historical snapshot.
The current package also supplies the small latent-RGB preview projection;
preview behavior therefore follows the installed code rather than changing an
older denoiser checkpoint tag.

Canter `0.7.0` adds inference-time `(text:weight)` emphasis for positive and
negative prompts. Prompt and learned-unconditional keys are normalized
separately, then both projected keys and values are interpolated toward or away
from the unconditional endpoint. Weight `1` is a no-op, weight `0` selects that
endpoint, and negative weights are rejected. The Python pipeline, Gradio
interface, and ComfyUI text encoder share this implementation.
The opt-in ER-SDE Brownian-tree solver uses a fixed mathematical clock domain,
allowing nested coarse and refined timestep grids to share one stochastic path.
The existing Euler--Maruyama, ER-SDE, and DPM++ 2M SDE names retain independent
per-step normals and their earlier seeded trajectories.

Canter `0.6.0` supports the extended denoiser topology used by `v0003`, with
text cross-attention in every prefix, middle, and suffix image layer. The same
package retains explicit support for the transition-only topology stored by
`v0001` and `v0002`. It also adds flow-matching DPM++ 2M SDE sampling with a
shared DPM++ eta/SDE noise-scale control. The UI uses eta `0.5` for DPM++ 2M
SDE and noise scale `1.0` for Euler-Maruyama and ER-SDE.

Canter `0.5.0` adds contrastive PDG. Its middle-skipped guidance path uses an
explicit negative prompt when supplied and learned unconditional text when the
negative prompt is blank or absent. The Python API, Gradio interface, and
ComfyUI nodes use the same branch-selection rule.

Canter `0.4.1` makes the dense backend apply to the bundled text encoder as
well as denoiser text attention. Dense text encoding now uses padded PyTorch
scaled dot-product attention and does not require the explicit CUDA
FlashAttention operator. The jagged backend retains its packed FlashAttention
implementation.

Canter `0.4.0` supports PyTorch 2.13 and adds public APIs for conditioning,
guidance configuration, guided velocity, schedules, solvers, the latent-RGB
projection, and DINAC VAE encoding. Euler-Maruyama and ER-SDE now use one
shared `sde_noise_multiplier` inference control. Gradio is provided by the
optional `webui` extra and is not an inference-package dependency.

Canter `0.3.0` added asynchronous native-resolution latent previews and the
responsive preview grid. The browser scales previews for display; final
decoded images remain full resolution.

| Release | Date | Weight storage | Status |
| --- | --- | --- | --- |
| [`v0003`](https://huggingface.co/data-archetype/canter/tree/v0003) | August 2026 | bfloat16 with float32 precision islands | Preview |
| [`v0002`](https://huggingface.co/data-archetype/canter/tree/v0002) | July 2026 | bfloat16 with float32 precision islands | Preview |
| [`v0001`](https://huggingface.co/data-archetype/canter/tree/v0001) | July 2026 | bfloat16 with float32 precision islands | Preview |

## Updating code

Refresh the moving `main` directory and use an editable installation:

```bash
hf download data-archetype/canter --revision main --local-dir canter
python -m pip install -e "./canter[webui]"
```

For a Git clone on `main`, use `git pull`. If the checkout was installed
without `-e`, reinstall it after updating.

## Selecting a checkpoint

Remote API and web UI loading can select any compatible checkpoint explicitly:

```python
pipe = CanterPipeline.from_pretrained(
    "data-archetype/canter",
    revision="v0001",
)
```

```bash
canter-web --model data-archetype/canter --revision v0001 --in-browser
```

These commands use the currently installed code and download only the selected
checkpoint artifacts. Running `app.py` inside a tagged standalone directory
instead uses the historical code bundled with that snapshot.

For exact reproduction, pin both the Canter package version and checkpoint
tag. Passing `revision="main"` explicitly selects the moving repository head
and is not an immutable checkpoint reference.

## v0003

Extended preview release with text cross-attention in all 30 image layers.

## v0002

Preview candidate exported from a 50/50 EMA blend of training steps 1,010,000
and 1,090,000.

## v0001

Initial preview release.