ChronoEdit Model (CRT AutoDL)

ChronoEdit is NVIDIA's image-editing model, and it's worth knowing what you're actually getting before you wire this node up. It's not a classic edit model like Qwen-Edit or InstructPix2Pix. Under the hood it's a Wan I2V video model - you generate a short clip of 2 or 26 frames and the last frame is your edited image. That sounds roundabout, but it's why the model got a genuinely good reception when it landed in late 2025: edits come out coherent and non-destructive, without the whole image degrading the way Qwen-Edit could. People asked for it in ComfyUI the same day it dropped.

This node is the "AutoDL" treatment: it downloads the model file itself and hands you a ready-to-go MODEL. No browsing HuggingFace, no figuring out which folder a 14GB file goes in.

How it works

The pack ships a registry of models with hard-coded filenames and HuggingFace URLs. When you first run this node it checks ComfyUI/models/diffusion_models for ChronoEdit_fp8_e4m3fn_scaled_VAI.safetensors; if it's missing, it pulls it from vantagewithai/ChronoEdit-fp8-scaled with a progress bar in your console, then loads it like any normal diffusion model. That fp8-e4m3fn weight is the reason this runs on consumer cards at all - it's a scaled-down 14B, the same quantization trick that made fp8 the default for big 2025-2026 video models. If the file's already there, the node just loads it; reruns are instant.

Because ChronoEdit is Wan-based, the file drops into diffusion_models and expects the Wan family around it - you'll see that reflected in the sibling nodes this pack ships.

The three inputs that matter

That's all there is, which is refreshing. The node has no required inputs, just three toggles:

• patch_cublaslinear (default off) - flips ComfyUI's fast cublas ops flag for the run. A modest speedup on some cards; leave it off unless you're chasing seconds.
• sage_attention (default auto) - the interesting one. It patches the model's attention with SageAttention, a fused flash-attention-style kernel that's a real speed/memory win on big DiTs.
• enable_fp16_accumulation (default on) - sets torch to accumulate matmuls in fp16 instead of fp32. Slightly faster, marginally less precise; the default is fine.

Where people get burned: sage_attention: auto does from sageattention import sageattn under the hood, and sageattention is not in this pack's requirements.txt. If you don't have it installed separately, auto throws an ImportError and the whole graph dies. If you've never installed it, set the dropdown to disabled and move on - you lose nothing except the speedup. On a modern NVIDIA card with pip install sageattention done, auto is worth it.

Wiring it up

The MODEL output feeds your KSampler like any other model. But ChronoEdit doesn't work in isolation - you need the whole set. This pack ships all of them under CRT/AutoDL/ChronoEdit, and each one auto-downloads its own piece: the ChronoEdit CLIP - WAN node (the umt5-xxl text encoder), ChronoEdit VAE (Wan's VAE), ChronoEdit Distill LoRA to patch the model for fewer steps, and ChronoEdit CLIP Vision if you're doing reference-image edits. Each downloads separately on first run, so your first workflow execution is a "wait, it's still downloading" moment. Be patient once; after that it's cached.

Install

It's just the pack. In ComfyUI Manager search CRT-Nodes and install, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes
pip install -r requirements.txt

Then restart ComfyUI. Note the requirements file is heavy (opencv, ultralytics, transformers, audio libs) because the pack does a lot of things - the AutoDL nodes themselves only really need huggingface_hub. If installing the whole list hurts, this node at minimum wants the pack registered; ComfyUI's bundled deps cover the core.