Self-Flow-diffusers

Diffusers-ready checkpoints for Self-Flow (Self-Supervised Flow Matching), converted from the official Self-Flow release for local/offline use.

This root folder is a model collection that contains:

• Self-Flow-XL-2-256

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

• pipeline.py (SelfFlowPipeline)
• transformer/transformer_selfflow.py and weights
• scheduler/scheduling_flow_match_selfflow.py (SelfFlowFlowMatchScheduler, SDE flow matching)
• scheduler/scheduler_config.json
• vae/ (stabilityai/sd-vae-ft-ema)

Each variant embeds English id2label in model_index.json, so class labels can be passed as ImageNet ids or English synonym strings.

Demo

Class-conditional sample (ImageNet class 207, golden retriever), Self-Flow-XL/2 at 256×256, 250 steps, CFG 3.5, seed 42.

Model Paths

Use paths relative to this root README:

Model	Resolution	Local path
Self-Flow-XL/2	256×256	./Self-Flow-XL-2-256

Recommended inference

Setting	Value
Resolution	256×256
Sampler	Self-Flow SDE flow matching
Steps	250
CFG scale	3.5
Guidance interval	(0.0, 0.7) when CFG > 1
Dtype	bfloat16 (recommended on Ampere+)
VAE	stabilityai/sd-vae-ft-ema

Inference Demo (Diffusers)

from pathlib import Path
import torch
from diffusers import DiffusionPipeline

model_dir = Path("./Self-Flow-XL-2-256").resolve()
device = "cuda" if torch.cuda.is_available() else "cpu"

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.to(device)

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

print(pipe.id2label[207])
print(pipe.get_label_ids("golden retriever"))  # [207]

image = pipe(
    class_labels="golden retriever",
    num_inference_steps=250,
    guidance_scale=3.5,
    generator=generator,
).images[0]
image.save("self_flow_xl_256_demo.png")