4-step photographic text-to-image (≈3 s per 1024² image on an RTX 5090)
and a companion instruction-editing model that takes up to three
reference images.
What this repo is: the Mage-Flow turbo and turbo-edit transformers, a
shared Qwen2.5-VL text encoder and VAE — weights only, not a retrain. The
settings below are the values these weights are actually run with day to
day.
Samples — mage-flow-turbo (txt2img)
Four steps, guidance 1.0. Photorealism is this model's strength.
prompt: a red apple on a rustic wooden table beside a window, soft daylight, fine skin texture and wood grain, studio photograph, sharp focus — 1024×1024, 4 steps, guidance 1.0, seed 12345
prompt: portrait of an older fisherman with a weathered face, natural window light, shallow depth of field, photorealistic — 1024×1024, 4 steps, guidance 1.0, seed 99
**prompt:** `a red apple on a rustic wooden table beside a window, soft daylight, fine skin texture and wood grain, studio photograph, sharp focus` — 1024×1024, 4 steps, guidance 1.0, seed 12345
**prompt:** `portrait of an older fisherman with a weathered face, natural window light, shallow depth of field, photorealistic` — 1024×1024, 4 steps, guidance 1.0, seed 99
prompt: a rain-slicked Tokyo street at night, neon signs reflecting in puddles, cinematic, 35mm photograph — 1024×1024, 4 steps, guidance 1.0, seed 1234
**prompt:** `a rain-slicked Tokyo street at night, neon signs reflecting in puddles, cinematic, 35mm photograph` — 1024×1024, 4 steps, guidance 1.0, seed 1234
Editing — mage-flow-edit-turbo (img2img)
Plain-language instructions against a reference image. Identity, pose,
lighting and camera are preserved; only what the instruction names changes.
reference
edit: put a brown flat cap on his head
edit: change the background to a sunny beach
**reference** — the seed-99 portrait above
**instruction:** `put a brown flat cap on his head` — 4 steps, seed 7 (≈88 s)
**instruction:** `change the background to a sunny beach with the ocean behind him` — 4 steps, seed 7 (≈60 s)
Both edits keep the subject essentially pixel-identical — same skin detail,
same jacket, same catchlights — while adding an object in one case and
replacing the entire environment in the other. Up to three reference
images may be supplied together (subject, style, context).
Four steps is correct. Both variants are step-distilled at guidance 1.0;
raising either is not the lever for quality.
Editing costs far more than generating — ≈60–90 s versus ≈3 s, because
the reference must be encoded through the VL tower before denoising begins.
That asymmetry is inherent to the architecture, not a misconfiguration.
Write edits as instructions, not descriptions — put a brown flat cap on his head, not a man wearing a brown flat cap.
torchvision is a hard dependency of the Mage-Flow pipeline (the VL
image preprocessing path) even though the rest of this stack does not need
it. Install the build matching your torch (torchvision==0.28.0 for torch
2.13.0+cu130) or loading fails with ModuleNotFoundError.
Three transformers ship here: base (30 steps, guidance 5.0 — the
undistilled model), turbo and turbo-edit (4 steps, guidance 1.0). The
samples on this card are from the two turbo variants; the base transformer
is included for anyone who wants the slower, higher-guidance path.
Each transformer directory is self-contained — download only the variant you
intend to run (8.2 GB each) plus the shared encoder, VAE and scheduler.
Layout
Components ship as separate directories under split/: one transformer
directory per variant (turbo, turbo-edit), plus the shared text encoder,
VAE and scheduler config. Loaders that accept a diffusers-style component tree
can consume this directly.
