caikybaldo999/zyi-flash

ZYI FLASH

A small text-to-image + instruction image-editing model at 256×256 — the next version of ZYI-256 Flash, fine-tuned in a single session on one NVIDIA RTX PRO 6000.

What's new vs ZYI-256 Flash

• Glyph branch: text in quotes (a sign that says "OPEN") is also read character by character by a frozen ByT5-small encoder, projected into the same cross-attention as flan-T5 — this is what lets the model spell words.
• Spell: quoted text is also spelled out in the caption ("OPEN" (O P E N)), in training and at inference (spell=True).
• 16 register tokens: a learned global "scratchpad" in the DiT sequence that helps counting and layout.
• Data emphasis: 3× more images with text, 2× more captions with counts, and a final high-quality annealing phase (top 35% by FLUX-Reason quality scores).
• Image editing (in-context source tokens + text/image CFG) is kept.

Generator	DiT, 1024d × 20 blocks, rectified flow, 2D RoPE, QK-norm, adaLN-single
Latent	DC-AE f32c32 — 256 px → 8×8×32 (64 tokens)
Text	flan-T5-base (frozen) + ByT5-small glyph encoder (frozen)
Training	+20,892 steps from caikybaldo999/zyi-256-flash, 75% text-to-image / 25% editing, 9.4 h
Data	1,000,000 images from FLUX-Reason-6M, 176,182 edit pairs from OmniEdit (MIT)

Usage

import importlib.util
from huggingface_hub import hf_hub_download
spec = importlib.util.spec_from_file_location("zyi_flash", hf_hub_download("caikybaldo999/zyi-flash", "zyi_flash.py"))
zf = importlib.util.module_from_spec(spec); spec.loader.exec_module(zf)
pipe = zf.ZyiPipeline("caikybaldo999/zyi-flash")
pipe('a shop sign that says "OPEN"', n=4, seed=0)                      # text -> image
from PIL import Image
pipe.edit(Image.open("photo.jpg"), "make it winter, add snow", seed=0)  # image editing

Requires torch, diffusers, transformers, safetensors, huggingface_hub. English prompts.

Editing — each row: source · edited · reference.

Limitations

256×256 output with an 8×8 latent grid: long text, small faces and fine details are still hard.