Qwen-Image-Edit-2511 ComfyUI template (image_qwen_image_edit_2511): fp8mixed versus bf16 file, two reference images and the 40/4 or 4/1 Lightning switch

Qwen-Image-Edit-2511 in ComfyUI: set up and run the official template Scope. This guide reads the official image_qwen_image_edit_2511.json in Comfy-Org/workflow_templates at commit 0cc0fba562efe9f68e48050c2b82ed6495050bc2 (SHA256 b2310393...9111, 59,841 bytes). The root graph has two LoadImage nodes, one wrapper node "Image Edit (Qwen-Image 2511)" (node 170) holding an embedded subgraph, and SaveImage. Every value was traced through links and the subgraph boundary. Nothing was run and no weights were downloaded. 1. Prerequisites - ComfyUI with subgraph support and the core nodes TextEncodeQwenImageEditPlus, FluxKontextImageScale, FluxKontextMultiReferenceLatentMethod, CFGNorm and ComfySwitchNode. All nodes are comfy-core, with saved versions 0.5.1 to 0.16.4 and frontend 1.45.15. These are saved strings, not a tested minimum. No custom node pack is needed. - Four files. Each loader node also declares its own download source: models/diffusion_models/qwen_image_edit_2511_fp8mixed.safetensors (Comfy-Org/Qwen-Image-Edit_ComfyUI; no size given in the file) models/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors (8.74 GB; declared from Comfy-Org/HunyuanVideo_1.5_repackaged, the same encoder file the HunyuanVideo 1.5 templates use) models/vae/qwen_image_vae.safetensors (242 MB; Comfy-Org/Qwen-Image_ComfyUI) models/loras/Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors (810.2 MB; lightx2v/Qwen-Image-Edit-2511-Lightning; only used with the Lightning switch on) - File-name mismatch: the template's Model Links note and storage tree list the diffusion model as qwen_image_edit_2511_bf16.safetensors (38.05 GB). UNETLoader 161 is set to qwen_image_edit_2511_fp8mixed.safetensors and declares that file's download. If you follow the note, select the bf16 file in the wrapper's unet_name; otherwise the loader looks for fp8mixed. 2. Inputs - Image 1 (LoadImage 41, example leather_sofa.png) is the image being edited. It goes to the wrapper's image1, through FluxKontextImageScale (160), and then to both text encoders as image1 and to VAEEncode (156) as the starting latent. There is no empty-latent size control: the output size follows image 1 after that scaling. - Image 2 (LoadImage 83, example texture_fur.png) goes unscaled to both encoders as image2. The wrapper also has an image3 input that is not connected; wire a third LoadImage to it if you need one. - Prompts: the positive TextEncodeQwenImageEditPlus (151) holds the example instruction to change the sofa's leather to the fur material in image 2; the negative (149) is empty. Both prompts are promoted to the wrapper. - Reference method: both conditionings pass through FluxKontextMultiReferenceLatentMethod (147 negative, 148 positive), saved as index_timestep_zero. A template note says these nodes are not needed with Comfy's own files but may be needed with files repackaged by others. 3. Settings that reach the sampler KSampler 169 uses euler with the simple scheduler and denoise 1. The model chain is UNETLoader 161, then ModelSamplingAuraFlow (145) at shift 3.1, then CFGNorm (152, saved 1 and false). Steps, CFG and the model input are linked to switches driven by one boolean, "Enable 4steps LoRA?" (node 168, promoted as enable_turbo_mode and saved off): - Off (saved default): 40 steps, CFG 4, model from CFGNorm without the LoRA. - On: 4 steps, CFG 1, model from CFGNorm through the Lightning LoRA (153) at strength 1. The KSampler's own saved widgets show 40 steps and CFG 3, but those inputs are linked, so the switch values (40/4 or 4/1) are what run; CFG 3 is never used. The seed is promoted on the wrapper: 677909188488042 with control randomize. Output goes through VAEDecode (158) to SaveImage 9 with prefix Qwen_Edit_2511. 4. How this relates to the published model-card records - The off route (40 steps, CFG 4) matches the diffusers multi-image recipe in "Qwen-Image-Edit-2511: multi-image pipeline settings and built-in LoRA effects" (40 steps, true_cfg_scale 4.0). The template's KSampler settings note lists the same Qwen values, plus a "Comfy" alternative of 20 steps at CFG 4.0. The saved switch uses 40. - The on route loads the 4-step bf16 LoRA described in "Qwen-Image-Edit-2511-Lightning: 4-step LoRAs and a fused FP8 model" and runs 4 steps, as that card says. The template also sets CFG 1, which the card itself does not state. The card's fused FP8 Lightning model is a different file from anything this template loads. 5. Run it 1) Put the four files in place, choosing either the fp8mixed model the loader names or the bf16 file the note names, and select it in unet_name. 2) Open the template. 3) Load the image to edit as image 1 and the reference as image 2. 4) Write an instruction that refers to the images by number, as the example does. 5) Leave Lightning off for 40 steps, or turn it on for 4 steps. 6) Run. Expect one image saved with prefix Qwen_Edit_2511 at image 1's scaled size. 6. Failure cases and checks - The diffusion model is missing because the note's bf16 name and the loader's fp8mixed name differ; select whichever file you downloaded. - Lightning switched on without the LoRA file in models/loras. - An output size that follows image 2: it will not, because size comes from image 1. - Expecting image3 to be used: it is unconnected by default. - Copying the KSampler's saved CFG 3 into another workflow: the effective CFG is 4, or 1 with Lightning on. - An unrepeatable result: the seed is set to randomize, so record it and set it to fixed. - Third-party repackaged models: keep the reference-method nodes, as the template note advises. 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 at this commit; that covers this analysis of the template, not the model weights, encoder, VAE, LoRA, example images or generated outputs, which keep their own terms.

{
  "citations": [
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/image_qwen_image_edit_2511.json",
      "note": "Exact pinned template; node 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)."
    }
  ]
}