README.md

---
license: mit
library_name: diffusers
pipeline_tag: text-to-image
tags:
- diffusers
- adm
- adm-g
- image-generation
- class-conditional
widget:
- output:
    url: ADM-G-512/demo.png
language:
- en
---

# BiliSakura/ADM-diffusers

Self-contained OpenAI ADM-G checkpoints for Hugging Face diffusers. **No external code repo is required** — each subfolder ships its own `pipeline.py`, component modules, and weights.

This repo is derived from the development bundle in [Visual-Generative-Foundation-Model-Collection](https://github.com/Bili-Sakura/Visual-Generative-Foundation-Model-Collection), but inference only needs:

- This model repo (`BiliSakura/ADM-diffusers`)
- PyPI `diffusers`, `torch`, `huggingface_hub`

This Hugging Face repo hosts **multiple self-contained checkpoints as subfolders**. Each subfolder includes its own `pipeline.py`, `model_index.json`, weights, and component code (`unet/`, `classifier/`, `scheduler/`).

## Available checkpoints

| Subfolder | Resolution | Guidance scale | OpenAI sources |
| --- | --- | ---: | --- |
| [`ADM-G-256/`](ADM-G-256/) | 256×256 | 1.0 | `256x256_diffusion.pt` + `256x256_classifier.pt` |
| [`ADM-G-512/`](ADM-G-512/) | 512×512 | 4.0 | `512x512_diffusion.pt` + `512x512_classifier.pt` |

Both resolutions use the **class-conditional** diffusion checkpoint plus the noisy classifier (not the 256 uncond variant).

## ImageNet class labels

Each variant keeps an `id2label` map directly in its own `model_index.json` (same style as DiT on the Hub). Runtime label resolution is English-only:

- `pipe.id2label` — inspect id → English label correspondence
- `pipe.labels` — reverse map (English synonym → id), sorted for browsing
- `pipe.get_label_ids("golden retriever")`
- `pipe(class_labels="golden retriever", ...)`

Chinese labels are still preserved in the main source repo under `src/labels/id2label_cn.json` for reference.

## Demo

![ADM-G-512 demo](ADM-G-512/demo.png)

Settings used for this demo image: `ADM-G-512`, `DDIMScheduler`, `num_inference_steps=50`, `guidance_scale=4.0`, `seed=42`, class `"golden retriever"`.

```python
from pathlib import Path
import torch
from diffusers import DDIMScheduler, DiffusionPipeline

model_dir = Path("./BiliSakura/ADM-diffusers/ADM-G-512")
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    local_files_only=True,
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)
pipe = pipe.to("cuda")
pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
class_id = pipe.get_label_ids("golden retriever")[0]
generator = torch.Generator(device="cuda").manual_seed(42)

out = pipe(
    class_labels=class_id,
    guidance_scale=4.0,
    num_inference_steps=50,
    generator=generator,
).images[0]
out

```

## Repo layout

```text
BiliSakura/ADM-diffusers/
├── README.md
├── ADM-G-256/
│   ├── pipeline.py
│   ├── model_index.json
│   ├── unet/
│   ├── classifier/
│   └── scheduler/
└── ADM-G-512/
    ├── pipeline.py
    ├── model_index.json
    ├── demo.png
    ├── unet/
    ├── classifier/
    └── scheduler/
```