ChronoEdit 14B ComfyUI template (image_chrono_edit_14B): five files, a 5-frame Wan I2V edit that keeps the last frame, and the Enable Distll LoRA 20/4 or 4/1 switch

ChronoEdit 14B image editing in ComfyUI: the official template Scope. This guide reads the official image_chrono_edit_14B.json in Comfy-Org/workflow_templates at commit 0cc0fba562efe9f68e48050c2b82ed6495050bc2 (SHA256 c8669763...5377, 54,830 bytes). The root holds 4 nodes and one subgraph, "Chrono Image Edit" (22 nodes), with one switch boolean. All nodes are active, and both switch states were traced statically. Nothing was run, and no model, image or linked page was downloaded. 1. Prerequisites - A ComfyUI build with the core nodes used here: WanImageToVideo, CLIPVisionEncode, ScaleROPE, ImageFromBatch, ComfySwitchNode and the primitives. They are comfy-core, saved at 0.3.67, with LoraLoaderModelOnly and the wrapper at 0.4.0. Those are saved version strings, not a tested minimum. No custom node pack is used. - Files, as the "For Local User" note (node 64) and the loaders declare them, with the note's sizes: - models/diffusion_models/chrono_edit_14B_fp16.safetensors (30.54 GB). UNETLoader 101 loads it with weight_dtype fp8_e4m3fn, so the fp16 file is cast to fp8 at load. - models/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors (6.27 GB), with CLIPLoader 108 type wan. - models/clip_vision/clip_vision_h.safetensors (1.18 GB), with CLIPVisionLoader 100. - models/vae/wan_2.1_vae.safetensors (242.1 MB), with VAELoader 99. - models/loras/chronoedit_distill_lora.safetensors (358.5 MB), with LoraLoaderModelOnly 111 at strength 1. It is used only when the switch is on. The links point to Comfy-Org/Wan_2.1_ComfyUI_repackaged and Comfy-Org/Wan_2.2_ComfyUI_Repackaged. The total is about 38 GB. - Per the note, ChronoEdit-14B is fine-tuned from Wan2.1-I2V-14B 720P, by NVIDIA and the University of Toronto. It treats image editing as a two-frame video generation task, and comes with the PBench-Edit benchmark. The project page was not followed. 2. The controls you edit The root holds LoadImage 52 (example image_chrono_edit_input_image.png), the wrapper 97 "Chrono Image Edit" and SaveImage 60 (prefix Chrono_Edit_14B). The wrapper promotes: - the prompt (CLIPTextEncode 105); - width and height (720 x 720 on WanImageToVideo 109); - the four model names (diffusion model, text encoder, VAE, CLIP vision); - Enable Distll LoRA (PrimitiveBoolean 123, saved off; its subgraph input is labelled enable_turbo_mode); - the KSampler seed (164026091171544, randomize). Inside the subgraph, the groups are "Step 1 - Load models (for local users)", "Step 2: Prompt", "Step3: Image size", "Distll LoRA settings" and "Original Settings". 3. How the edit is made - The input image goes to CLIPVisionEncode 104 (crop none, clip_vision_h) and to the start_image of WanImageToVideo 109. Node 109 also takes the positive prompt 105, the negative prompt 106, the VAE and the CLIP vision output. - WanImageToVideo 109 is set to 720 x 720, length 5, batch 1. After sampling, VAEDecode 102 decodes the video, and ImageFromBatch 103 (saved 4, 1) keeps one image, at index 4: the fifth and last frame. That is the edited image, which the subgraph outputs to SaveImage 60. - The template has no separate temporal-reasoning control. Length 5 is a saved value, and the reason for 5 frames is not stated in the file. 4. Sampling and the switch - Model path: UNETLoader 101 feeds ModelSamplingSD3 107 (shift 5), then ScaleROPE 98 (saved 1, 0, 1, 0, 7, 0; the names are not serialized in the file). Switch (Model) 117 then picks the ScaleROPE output when off, or LoraLoaderModelOnly 111 when on; that LoRA is applied on top of the ScaleROPE output. So ScaleROPE and shift 5 apply on both routes. - Switch (Step) 118 gives 20 (PrimitiveInt 113) off or 4 (116) on. Switch (CFG) 119 gives 4 (PrimitiveFloat 122) off or 1 (121) on. All three switches read Enable Distll LoRA. - KSampler 110 uses uni_pc, scheduler simple, denoise 1, and seed randomize. Its steps and CFG are linked from the switches; its own saved 20 and 4 match the off route. - In summary: saved off, 20 steps at CFG 4 without the LoRA; on, 4 steps at CFG 1 with chronoedit_distill_lora. 5. Prompts The saved positive prompt describes the whole target scene: a facial cleansing foam bottle floating among fluffy bubbles on a light pink background, in a professional advertising style. It is not a short edit instruction. The negative prompt (106) is the usual Wan Chinese quality list. The example image was not viewed, so how it relates to the prompt is not established. 6. Run it 1) Put the five files in the folders above. You can skip the LoRA if you keep the switch off. 2) Open the template and load your image in LoadImage 52. 3) On the wrapper, write the prompt for the result you want, and set width and height. 4) Leave Enable Distll LoRA off for 20 steps at CFG 4, or turn it on for 4 steps at CFG 1. 5) Run. Expect one image under Chrono_Edit_14B, taken from the last of the 5 generated frames. 7. Failure cases and checks - Missing LoRA with the switch on: the switch routes the model through LoraLoaderModelOnly 111, so the LoRA file must exist. - Memory: the fp16 checkpoint is 30.54 GB on disk, though it is loaded as fp8. The file gives no VRAM or timing figures. - Size changes: width and height also scale the generated frames, and the file states no size rules. Keep the input image's aspect ratio close to the output size. - Wrong frame: changing length on node 109 without changing ImageFromBatch 103's index (4) would save a frame other than the last. - Edits that seem ignored: try describing the target result, as the example prompt does. Whether short instructions work as well was not tested here. - Saved versions: if nodes are missing, update ComfyUI. 0.3.67 and 0.4.0 are saved versions, not known minimums. 8. Related templates The one-switch pattern matches the lane's Qwen-Image-Edit-2511 guide (Enable 4steps LoRA: 40/4 or 4/1) and the LoRA switch comparison, which does not cover ChronoEdit. The pipeline uses WanImageToVideo with CLIP vision, the same core node family as the Wan I2V templates, which fits the note's statement that ChronoEdit is fine-tuned from Wan2.1-I2V-14B. Evidence and rights. Evidence status: source_inspected, not_run, unreviewed. This is shared-founder desk curation by the tutorial lane. The workflow repository declares the MIT licence; that covers this analysis of the template, not the ChronoEdit weights, LoRA, encoders, example image or outputs, which keep their own terms.

{
  "citations": [
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/image_chrono_edit_14B.json",
      "note": "Exact pinned template; node, link and subgraph byte ranges and excerpt hashes are in the source pins."
    },
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/LICENSE",
      "note": "Repository MIT licence at the same commit (existing native source)."
    }
  ]
}