Introduction

NVIDIA's ChronoEdit-14B treats an image edit as a tiny five-frame video. That makes its edits pleasingly physical, but the default recipe is slow: 50 steps with CFG 5, or 100 network evaluations per edit. A new open checkpoint, ChronoEdit-14B-DMD-4step from researcher Chenming Wu, gets the same job done in 4 evaluations with no guidance.

It landed on Hugging Face at 23:28 UTC on October 31 (07:28 MYT, November 1). It's open weights under Apache-2.0, like the base model, and it runs today in Python through diffusers and NVIDIA's ChronoEdit repo. It's a drop-in transformer swap: the VAE, text encoder and image encoder still come from nvidia/ChronoEdit-14B-Diffusers.

What Shipped

  • Model: ChronoEdit-14B-DMD-4step (ChenmingWu/ChronoEdit-14B-DMD-4step)
  • Modality: instruction-based image editing (image + text → edited image)
  • Method: distribution matching distillation (DMD) in the DMD2 / Self-Forcing style, with a trainable fake score
  • Steps: 4, at sigmas 1.0, 0.75, 0.5 and 0.25 with shift 5, guidance 1.0 (no CFG)
  • Tested setting: 720p (1280×720 area), 5 frames, keep the last frame
  • License: Apache-2.0
  • Format: diffusers transformer only, split into 7 safetensors shards

The Numbers

The card compares three setups at 720p. All of them are scored by open-source judges, so don't line these numbers up against papers judged by GPT-4.

  • ChronoEdit-14B, 50 steps, CFG 5 (100 NFE): GEdit overall 5.644, ImgEdit 3.487
  • Official 8-step distill LoRA (8 NFE): GEdit overall 3.637, ImgEdit 3.444
  • This DMD model (4 NFE): GEdit overall 5.772, ImgEdit 3.306

So on GEdit-Bench-EN it lands on par with the full 50-step model, and well ahead of NVIDIA's own 8-step LoRA on semantic consistency (5.860 vs 3.621). ImgEdit dips a little: the paired bootstrap puts it at −0.17 against the 50-step model. The author also flags that the checkpoint was picked on a quarter of GEdit. On the held-out three quarters it still scores 5.469 against 5.242 for the 50-step model.

How it was trained

  • The generator started from ChronoEdit-14B with the official 8-step LoRA merged in, then got full fine-tuning.
  • The real score was ChronoEdit-14B at CFG 5. A trainable copy learned the fake score, with 4 critic updates per generator update.
  • A quarter of the noise draws sit at sigma 0.98–1.0, where ChronoEdit's guidance actually decides the edit.
  • Training ran 200 iterations at batch 7 on 7 GPUs, using about 9.5k source images and instructions from OmniEdit, HQ-Edit and UltraEdit. The released weights are the EMA.

Run It

Set up NVIDIA's ChronoEdit diffusers pipeline as its README describes, passing transformer=transformer when you build ChronoEditPipeline from nvidia/ChronoEdit-14B-Diffusers. Then swap in the 4-step sampler:

import torch
from diffusers import FlowMatchEulerDiscreteScheduler
from chronoedit_diffusers.transformer_chronoedit import ChronoEditTransformer3DModel

transformer = ChronoEditTransformer3DModel.from_pretrained(
    "ChenmingWu/ChronoEdit-14B-DMD-4step", subfolder="transformer", torch_dtype=torch.bfloat16)

pipe.scheduler = FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, shift=5.0, stochastic_sampling=True)
set_timesteps = pipe.scheduler.set_timesteps
pipe.scheduler.set_timesteps = lambda num_inference_steps=None, device=None, **kw: set_timesteps(
    sigmas=[1.0, 0.75, 0.5, 0.25], device=device)

frames = pipe(image=image, prompt=instruction, height=h, width=w, num_frames=5,
              num_inference_steps=4, guidance_scale=1.0, enable_temporal_reasoning=False).frames[0]
edited = frames[-1]

A few practical notes:

  • It's still a 14B transformer. NVIDIA's README puts the base pipeline at about 34 GB of GPU memory with offloading on, so plan for a big card or aggressive offload.
  • Only the 720p, 5-frame, 4-step recipe was tested. Treat other resolutions and samplers as experiments.
  • There's no ComfyUI single-file conversion yet. Today this is a diffusers-first release.

Limits and License Fine Print

The author is refreshingly candid about the weak spots:

  • It still over-edits sometimes, for example recolouring objects that should stay put or changing a whole subject's material.
  • Text rendering is unreliable.
  • Object removal is weak, just as in the base model.

The weights are Apache-2.0, but about a third of the training prompts come from HQ-Edit, which is CC BY-NC 4.0. The card asks you to check whether that matters before commercial use, and I'd take that advice seriously.

Conclusion

Going from 100 evaluations to 4 changes how an editor feels. With ChronoEdit's physical, video-style editing, that's the difference between waiting for each edit and trying ideas back to back. It's a solo research drop with zero likes so far, but it's well measured, honestly caveated and Apache-licensed. If you already run ChronoEdit in diffusers, it's an easy swap to try today.

Original Source

Four steps, endless edits. Have fun!

—Aurelia ♡