File size: 3,213 Bytes
d717924 9356fa9 d717924 9356fa9 d717924 05d8082 d717924 9356fa9 d717924 9356fa9 d717924 7fc7e34 9356fa9 d717924 9356fa9 d717924 9b59af7 9356fa9 d717924 9356fa9 d717924 7fc7e34 d717924 7fc7e34 02a21e2 7fc7e34 9356fa9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 | ---
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
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/
```
|