Instructions to use ChenmingWu/ChronoEdit-14B-DMD-4step with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use ChenmingWu/ChronoEdit-14B-DMD-4step with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline from diffusers.utils import load_image # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("ChenmingWu/ChronoEdit-14B-DMD-4step", dtype=torch.bfloat16, device_map="cuda") prompt = "Turn this cat into a dog" input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png") image = pipe(image=input_image, prompt=prompt).images[0] - Notebooks
- Google Colab
- Kaggle
import torch
from diffusers import DiffusionPipeline
from diffusers.utils import load_image
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("ChenmingWu/ChronoEdit-14B-DMD-4step", dtype=torch.bfloat16, device_map="cuda")
prompt = "Turn this cat into a dog"
input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png")
image = pipe(image=input_image, prompt=prompt).images[0]ChronoEdit-14B, 4-step DMD transformer
A guidance-free 4-step transformer for ChronoEdit-14B, obtained by
distribution matching distillation (DMD) from ChronoEdit-14B itself. It replaces only the transformer: the VAE, text encoder,
tokenizer and image encoder are those of nvidia/ChronoEdit-14B-Diffusers.
One edit takes 4 network evaluations (no classifier-free guidance), against 100 for ChronoEdit's default of 50 steps with CFG 5. The current main release uses anchoring strength 0.3 and prioritizes ImgEdit. Compared with the initial release, it has a higher ImgEdit point estimate and a lower GEdit score; see the full comparison below.
Updated 2026-11-05: main now contains the anchored EMA checkpoint at iteration 200. The initial unanchored release
is preserved at revision cfg5-no-anchor-step200-e72115b9.
Results
Setting for every row: 720p (area 1280x720, sides multiple of 16), 5 frames, last frame kept, seed 0, the benchmarks' raw instructions (no prompt enhancer). Judges are open-source models, so the absolute numbers are not comparable with papers that use GPT-4.1 / GPT-4o:
- GEdit-Bench-EN (606 items): official VIEScore protocol with Qwen2.5-VL-72B-Instruct-AWQ; SC = semantic consistency, PQ = perceptual quality, O = overall.
- ImgEdit basic set (737 items): ImgEdit-Judge, mean of the 9 edit types. We report both the official parser and a
lenient parser that also accepts scored lines such as
Score 1: 2 (Fair); both use the same judge outputs.
| Model | NFE | GEdit SC | GEdit PQ | GEdit O | ImgEdit (official / lenient) | |---|---|---|---|---|---|---| | ChronoEdit-14B, 50 steps, CFG 5 | 100 | 5.803 | 7.269 | 5.644 | 3.487 / 3.466 | | ChronoEdit-14B + official 8-step distillation LoRA | 8 | 3.621 | 7.589 | 3.637 | 3.444 / 3.457 | | Initial 4-step release, no anchoring | 4 | 5.860 | 7.468 | 5.772 | 3.306 / 3.324 | | Current main: anchor 0.3, iteration 200 | 4 | 5.551 | 7.469 | 5.465 | 3.400 / 3.379 |
- The current release trades GEdit performance for a higher ImgEdit point estimate. Relative to the initial release, the table's GEdit O changes by -0.307 and official ImgEdit by +0.094. The ImgEdit increase is not statistically established in the paired analysis below.
- Paired bootstrap, 95% CI, per-item means (different aggregation from the table): against the 50-step teacher, GEdit O -0.120 [-0.362, +0.125] over 606 pairs; ImgEdit -0.103 [-0.195, -0.015] over 737 pairs using the lenient parser. Against the initial unanchored release: GEdit O -0.338 [-0.518, -0.162]; ImgEdit +0.062 [-0.003, +0.131].
- All 606 GEdit images and judgements, and all 737 ImgEdit images and judgements, are present and checked. The official ImgEdit parser scores 651/737 current-release outputs; the lenient parser scores all 737. Parser omissions are not missing judgements.
- This checkpoint was selected on a fixed quarter of GEdit-Bench-EN (every 4th English item, 152 examples) from complete exports at iterations 100, 200 and 225. The planned 300-iteration run was interrupted while exporting iteration 250; the released iteration-200 export is complete. On the other 454 GEdit items its per-item O is 5.097, against 5.242 for the 50-step teacher and 5.469 for the initial release. Those 454 examples were excluded from checkpoint selection within the run; the full benchmarks subsequently informed the choice of released variant, so they are not untouched tests.
- For the initial release, pass
revision="cfg5-no-anchor-step200-e72115b9"tofrom_pretrained.
Usage
The model needs the chronoedit_diffusers package from the ChronoEdit repository
(ChronoEditPipeline, ChronoEditTransformer3DModel). Set up the pipeline as in ChronoEdit's own diffusers inference
instructions, then swap in this transformer and the 4-step sampler the model was trained with:
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 = ChronoEditPipeline.from_pretrained("nvidia/ChronoEdit-14B-Diffusers", transformer=transformer, ...) # as in ChronoEdit's README
# 4 steps at sigma = 1, 0.9375, 0.8333, 0.625 (shift 5 of 1, .75, .5, .25); after each step the predicted x0 is re-noised
# with fresh noise to the next level (LCM-style sampling)
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]
Training
DMD with a trainable fake score, in the style of DMD2 / Self-Forcing:
- Generator (this model): initialised from ChronoEdit-14B with its official 8-step distillation LoRA merged; 4-step sampler as above, no guidance; full fine-tuning.
- Real score: ChronoEdit-14B with CFG 5 and an empty negative prompt.
- Anchoring: add
0.5 * 0.3 * mean((x0 - x0_init)**2)to the generator loss at the selected rollout state. The frozen anchor is the generator's initialization (ChronoEdit-14B with the official distillation LoRA merged). - Fake score: a trainable copy of ChronoEdit-14B, trained by flow matching on the generator's outputs, 4 updates per generator update; 20 critic-only warm-up iterations.
- Noise levels for both scores: 25% drawn uniformly from [0.98, 1], where ChronoEdit's guidance decides the edit (its x0 prediction from a noised copy of the source shows the guided edit only at sigma >= ~0.98); the rest shift-5 levels in [0.02, 0.98].
- AdamW with beta1 = 0, learning rate 5e-7 (generator) and 2e-6 (fake score), 200 iterations of batch 7 (7 GPUs), 720p, 5 frames; the released weights are the EMA of the generator.
- Conditions: about 9.5k source images with their editing instructions sampled from OmniEdit-Filtered-1.2M, HQ-Edit and UltraEdit. The DMD loss uses no target images.
Limitations
- This anchored release does not dominate the initial release: ImgEdit's point estimate is higher, while GEdit is lower.
- Over-edits remain in some cases (colour changes of objects that should stay, whole-subject material changes); text rendering is unreliable; object removal is weak, as in the base model.
- Only 720p with 5 frames and the sampler above were tested.
- The checkpoint was selected on a quarter of GEdit-Bench-EN. Benchmark results also informed variant selection, as disclosed above; they should not be treated as independent confirmation on an unseen test set.
License
Apache-2.0, as ChronoEdit-14B (NVIDIA) on which it is built. Note on training data: about a third of the source images and instructions come from HQ-Edit, which is released under CC BY-NC 4.0 (OmniEdit-Filtered-1.2M: MIT; UltraEdit: CC BY 4.0); check whether this matters for your use before any commercial use.
Acknowledgements
Built on ChronoEdit (NVIDIA) and Wan2.1.
- Downloads last month
- -
Model tree for ChenmingWu/ChronoEdit-14B-DMD-4step
Base model
nvidia/ChronoEdit-14B-Diffusers