leffff/diffusion-alignment-workshop

Diffusion Alignment Workshop dataset

Offline data for a hands-on workshop on aligning a diffusion model with rewards (SFT, RWR, DPO, online DPO, GRPO, DiffusionNFT, on-policy distillation). The model is Sana-600M-512px and the task is text rendering: every prompt asks for a short string to be written in the image.

For every prompt there are 5 samples of the model, each scored with four rewards. With it, a Colab student can run the offline methods without sampling or scoring anything, and the online methods without the 2.6B Gemma-2 text encoder. It is the alignment companion of leffff/diffusion-distillation-workshop (same model, same embedding format).

split	prompts	samples
train	2048	5 per prompt
eval	64	5 per prompt

Prompts come from leffff/Diffusion-Reward-Modeling-for-Text-Rendering-Dataset (train / test splits), filtered to:

• exactly one quoted target, which must be written in the image: a rendering keyword next to the quote (written, spelled, the word, sign, label, a mug that says, ...), and not speech or an attribution (he says "...", "...", says Stan Lee, written by);
• short ASCII targets that a 600M model at 512 px can plausibly render: at most 3 words and 16 characters;
• at most 2 prompts per target string (there are 219 "Happy Birthday" prompts in the source).

Columns (one row per prompt)

column	content
prompt, text	the prompt and the string that must appear in the image
source_id	the row id in the source dataset
seeds	5 seeds, torch.Generator("cuda").manual_seed(seed) for the initial noise of each sample
images	5 decoded 512×512 JPEGs (quality 95)
latents	float16 bytes, shape (5, 32, 16, 16). DC-AE latents in model space (the x₀ the transformer sees). Decode with vae.decode(latent / 0.41407)
ocr	5 × OCR-Levenshtein reward in [0, 1]: 1 - min(d, len(t)) / len(t), where d is the fewest character edits that make the target t appear somewhere in the transcription (case, spaces and punctuation ignored)
ocr_read	5 × what the OCR model (Qwen3-VL-2B-Instruct, greedy) read
pickscore	5 × PickScore (logit_scale · cos, ≈ 15–25)
aesthetic	5 × LAION aesthetic predictor v2 on CLIP ViT-L/14 (weights, ≈ 1–10)
clip	5 × CLIP score, 100 · cos(image, prompt) with openai/clip-vit-large-patch14
prompt_embeds	float16 bytes, shape (128, 2304). Gemma-2 hidden states from Sana's encode_prompt with the complex human instruction, max_sequence_length=128
prompt_attention_mask	uint8 bytes, shape (128,)

null_prompt_embeds.safetensors holds the unconditional (empty negative prompt) embedding and mask for classifier-free guidance. meta.json holds all generation parameters and the reward definitions.

Generation

SanaPipeline default sampler (flow DPM-Solver), 20 steps, CFG 4.5, 512×512, transformer in fp16, Gemma-2 in bf16, DC-AE in fp32, on H100s. The rewards are computed on the decoded fp32 images before JPEG encoding. Code: generate_dataset.py and rewards.py of the workshop.

Loading

import numpy as np, torch, pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

path = hf_hub_download("leffff/diffusion-alignment-workshop", "data/train-00000-of-00008.parquet", repo_type="dataset")
t = pq.read_table(path, columns=["prompt", "text", "latents", "ocr", "prompt_embeds", "prompt_attention_mask"]).to_pydict()
latents = torch.from_numpy(np.stack([np.frombuffer(b, np.float16).reshape(5, 32, 16, 16) for b in t["latents"]]))
embeds = torch.from_numpy(np.stack([np.frombuffer(b, np.float16).reshape(128, 2304) for b in t["prompt_embeds"]]))
ocr = torch.tensor(t["ocr"])        # (n, 5)

License

Apache-2.0 for the images, latents and rewards generated by the apache-2.0 Sana-600M model. The prompts come from the MIT-licensed leffff/Diffusion-Reward-Modeling-for-Text-Rendering-Dataset.