What if your image editor could think before it edits literally reasoning through every pixel like a mini simulation of reality? That's exactly what NVIDIA's ChronoEdit-14B does. It doesn't just tweak images, it imagines how an edit would naturally unfold over time, like frames in a short video.
ChronoEdit-14B is a physics aware image editing and action conditioned world simulation model that brings temporal reasoning into visual editing. Built by NVIDIA under the ChronoEdit family of multimodal foundation models, it's powered by a massive 14-billion-parameter video generative backbone and is licensed for commercial use under both the NVIDIA Open Model License Agreement and Apache License 2.0.
 |
| Chrono edit Pipeline (Ref- Research Paper) |
The magic lies in its two-stage inference pipeline: first, a video reasoning stage denoises latent trajectories using temporal reasoning tokens,these tokens act like the model's thought process, ensuring edits stay physically plausible. Then, in the in-context editing stage, the model prunes these reasoning tokens to generate the final refined image efficiently.
In simple terms, ChronoEdit reframes image editing as a video generation task,using the input and edited images as start and end frames. You can find more detailed insights by accessing their research paper. This lets the model maintain perfect temporal consistency and even visualize how it "thinks" through an edit by denoising those reasoning tokens into a short video.
Installation
1. Setup ComfyUI if you haven't done yet. Update it from the Manager if you are an old user.
2. Download Chronoedit-models (any of them):
(a) Chronoedit FP16 For Image edit (chrono_edit_14B_fp16.safetensors by official ComfyUI)
(b) Chronoedit FP8 For Video Edit (Wan2_1-14B-I2V_ChronoEdit_fp8_scaled_KJ.safetensors Optimzed by Kijai)
The kijai's workflow uses Wan2.1, you need to have the Kijai's custom node ComfyUI Wan Video Wrapper installed.
(c) Chronoedit FP16 (Wan2_1-I2V-14B_ChronoEdit_fp16.safetensors Optimzed by Kijai)
Save it into ComfyUI/models/diffusion_models folder.
You can know more about optimized models from our quantized model tutorial.
3. Download Distill lora (any of them):
(a) Chronoedit distill lora (Wan_2_1_I2V_14B_ChronoEdit_distill_lora_rank32.safetensors by Kijai) or
(b) Chrono edit distill lora (chronoedit_distill_lora.safetensors by nvdia )
and save it into ComfyUI/models/loras folder.
5. Restart and refresh comfyui.
Workflow
1. Download the workflow by searching Chrono Edit on templaes section of ComfyUI. Another workflow is by Kijai. The workflow uses Wan2.1, you need to have the Kijai's ComfyUI Wan Video Wrapper installed. The workflow can be found inside ComfyUI/custom_nodes/ComfyUI-WanVideoWrapper/example_workflows folder.
2. Drag and drop into ComfyUI.
(a) Add your image to edit into Load image node.
(b) Upload other relevant models (Chrono Edit, lora, Wan 2.1 I2V, vae etc)
(c) Set KSampler Settings:
CFG-1.0
Steps-8 (with lora enabled)
Shift-2
(d) Add prompts- You need to use long and detailed prompts to generate good results. You can use any LLM (Chatgpt/Gemini/Deepseek/Qwen etc)to enhance your short prompts into more descriptive one. Short prompts often give you bad results.
(e) Hit Run to execute the workflow.
