This repository provides a minimal diffusion model for generating MNIST digits (0–9) from text prompts. It uses a tiny UNet with FiLM conditioning and Classifier-Free Guidance (CFG).

Quickstart (dynamic import)

python

from huggingface_hub import snapshot_download
import importlib.util, sys, os

local_dir = snapshot_download("starkdv123/mnist-28px-text2img")
pipe_path = os.path.join(local_dir, "pipeline_mnist.py")
spec = importlib.util.spec_from_file_location("pipeline_mnist", pipe_path)
mod = importlib.util.module_from_spec(spec)
sys.modules[spec.name] = mod
spec.loader.exec_module(mod)

Pipe = getattr(mod, "MNISTTextToImagePipeline")
pipe = Pipe.from_pretrained(local_dir)
img = pipe("seven", num_inference_steps=120, guidance_scale=2.5).images[0]
img.save("seven.png")

Alternative (sys.path)

python

from huggingface_hub import snapshot_download
import sys

local_dir = snapshot_download("starkdv123/mnist-28px-text2img")
sys.path.append(local_dir)
from pipeline_mnist import MNISTTextToImagePipeline as Pipe
pipe = Pipe.from_pretrained(local_dir)

Alternative (Diffusers custom_pipeline)

python

from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained("starkdv123/mnist-28px-text2img", custom_pipeline="pipeline_mnist")
# Note: requires diffusers>=0.30 and may still be sensitive to custom components.

Training details

Dataset: MNIST train split
Image size: 28×28 (grayscale)
Normalizer: (x - 0.5) / 0.5 → network predicts noise in [-1,1]
Timesteps: T=500, betas 1e-05→0.02
UNet base channels: 32; cond dim: 64; time dim: 128
CFG: drop‑cond p=0.1, null class id=10
Optimizer: AdamW, lr=0.001, batch=128, epochs=10

Files included

model.safetensors — trained UNet weights
config.json — model hyperparameters
scheduler_config.json — noise schedule
pipeline_mnist.py — custom pipeline definition
model_index.json — metadata for Diffusers
samples/grid_0_9.png — final 0–9 grid
samples/grid_e*.png — per‑epoch grids (training progress)

Training progress (epoch grids)