DeCo Diffusers

DeCo-diffusers

Diffusers-ready checkpoints for DeCo (Decoupled Conditioning), converted for local/offline use.

This root folder is a model collection that contains:

• DeCo-XL-16-256
• DeCo-XL-16-512
• DeCo-XXL-16-512-t2i (text-to-image; requires Qwen/Qwen3-1.7B text encoder)

Each subfolder is a self-contained Diffusers model repo with:

• pipeline.py
• transformer/transformer_deco.py
• scheduler/scheduling_deco_flow_match_euler_discrete.py
• transformer/diffusion_pytorch_model.safetensors
• vae/autoencoder_deco.py

Each variant embeds English id2label directly in model_index.json (DiT-style), so class labels can be passed as ImageNet ids or English synonym strings.

• pipe.id2label — id → English label (comma-separated synonyms)
• pipe.get_label_ids("golden retriever") — English label → id

Demo

DeCo-XL-16-512 demo

Class-conditional sample (ImageNet class 207, golden retriever), DeCo-XL/16 at 512×512, 100 steps, CFG 5.0, seed 42.

Model Paths

Use paths relative to this root README:

Model	Resolution	Source checkpoint	Local path
DeCo-XL/16	256×256	imagenet256_epoch800.ckpt (EMA)	./DeCo-XL-16-256
DeCo-XL/16	512×512	imagenet512_epoch340.ckpt (EMA)	./DeCo-XL-16-512
DeCo-XXL/16	512×512 t2i	t2i_DeCo.ckpt (EMA)	./DeCo-XXL-16-512-t2i

Inference Demo (Diffusers)

1) Load a local subfolder checkpoint

import torch
from diffusers import DiffusionPipeline

model_path = "./DeCo-XL-16-512"  # change to ./DeCo-XL-16-256 for 256px
device = "cuda" if torch.cuda.is_available() else "cpu"

pipe = DiffusionPipeline.from_pretrained(
    model_path,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to(device)

generator = torch.Generator(device=device).manual_seed(42)

# ImageNet class example: 207 = golden retriever
print(pipe.id2label[207])
print(pipe.get_label_ids("golden retriever"))  # [207]

result = pipe(
    class_labels="golden retriever",
    num_inference_steps=100,
    guidance_scale=5.0,  # use 3.2 for DeCo-XL-16-256
    generator=generator,
)

image = result.images[0]
image.save("deco_xl_512_demo.png")

2) Quick variant switch (256 model)

model_path = "./DeCo-XL-16-256"

pipe = DiffusionPipeline.from_pretrained(model_path, trust_remote_code=True).to(device)
image = pipe(
    class_labels=207,
    num_inference_steps=100,
    guidance_scale=3.2,
    generator=generator,
).images[0]
image.save("deco_xl_256_demo.png")

Integer class ids, batched labels, and optional batch_size for repeating a single label are also supported.

3) Text-to-image (DeCo-XXL-16-512-t2i / t2i_DeCo.ckpt)

Use the AdamLM scheduler defaults from official DeCo (not the c2i 100-step / CFG 5.0 recipe):

import torch
from diffusers import DiffusionPipeline

model_path = "./DeCo-XXL-16-512-t2i"
device = "cuda" if torch.cuda.is_available() else "cpu"

pipe = DiffusionPipeline.from_pretrained(
    model_path,
    trust_remote_code=True,
    custom_pipeline=f"{model_path}/pipeline.py",
    torch_dtype=torch.bfloat16,
).to(device)

# Bundled ./text_encoder (Qwen3-1.7B weights + tokenizer). Pipeline loads both from that folder.
# Denoiser runs in float32 during __call__ (matches official GenEval predict).

image = pipe(
    prompt="a golden retriever playing in the snow, high quality photograph",
    negative_prompt="Unrealistic, JPEG artifacts.",
    num_inference_steps=25,
    guidance_scale=4.0,
    timeshift=3.0,
    generator=torch.Generator(device="cpu").manual_seed(42),
).images[0]
image.save("deco_t2i_demo.png")