How do I keep the same AI character across images?

1. 01 / processReference

   Keep one approved original

2. 02 / processScene edit

   Change the setting, name fixed traits

3. 03 / processIdentity check

   Compare face, shape and outfit

4. 04 / processExport

   Save PNG plus reproduction log

Start from the same approved reference image and ask an image-editing model to change the scene. Describing the character again in a fresh text-to-image prompt can produce a recognisable type of character without preserving its identity. A seed is useful for recording a comparison; it is not a character identifier.

This guide uses Qwen-Image-Edit-2511 locally and Midjourney's paid Edit Model in the browser. Both accept visual references. Neither guarantees identical faces, clothes or proportions. These are documentation-based instructions: we have checked the sources and inspected the local template, but have not generated the proposed scenes.

Prepare one reference and a short identity sheet

Use an image you own or have permission to edit. For a fictional character, pick a clear, single-subject image with visible distinguishing features. For a person, obtain permission before uploading their photo. Keep the original file unchanged.

Write down what must survive: head shape, eye colour, hairstyle, identifying marks and clothing. Our suggested example is a cream-coloured robot with a rectangular face, two dark circular eyes and an orange jacket. That is a proposed task using the site's existing input, not a generated consistency demonstration.

Separate fixed traits from scene instructions. Choose three simple scenes—standing in a kitchen, sitting on a park bench, beside a shop window—and make each directly from the original reference. Repeatedly editing the previous output can accumulate unwanted changes.

Choose your route before downloading

Local Qwen editing: no hosted generation fee, but it needs ComfyUI, large model files, storage, electricity and a suitable execution environment. The Qwen model card lists Apache 2.0 and describes improved consistency; those are provider statements, not our measured results. Check the licence of every additional component separately.

Paid Midjourney: uploads references and runs on the provider's hardware. It needs an account and a subscription; consider the service's visibility settings before uploading private images. Plan documentation specifies recurring subscriptions and GPU-time allowances. Payment buys access and convenience, not an identity guarantee.

If local fit is uncertain, use the PC requirements answer first. We have no verified minimum RAM/VRAM or speed measurement for this recipe.

Free local route: prepare the exact workflow

1. Install or update ComfyUI using its official requirements and installation links. Save the actual version you use. A working installation alone does not mean this model fits your hardware.
2. Obtain the official Qwen 2511 template as JSON using GitHub's raw/download control, then open it in ComfyUI. The native tutorial also links the template library. Use this named template, not an unrelated community graph.
3. From the native tutorial's model links, prepare qwen_image_edit_2511_bf16.safetensors in ComfyUI/models/diffusion_models/, qwen_2.5_vl_7b_fp8_scaled.safetensors in models/text_encoders/, and qwen_image_vae.safetensors in models/vae/. These are separate components, not interchangeable checkpoints. Check file sizes and available resources before downloading.
4. Select those exact filenames in the editing subgraph's unet_name, clip_name and vae_name controls. The JSON inspected on 11 October 2026 defaults to an fp8mixed diffusion filename while its notes/tutorial list BF16. Explicitly selecting the downloaded BF16 file prevents that mismatch. Leave enable_turbo_mode false: this route uses the ordinary model, not the optional Lightning adaptation.
5. Replace the first LoadImage input with your reference. Disconnect the second sample image from the subgraph's image2 input; leave image3 disconnected. The supplied material-replacement sample contains a sofa and fur texture. Keeping either sample as an extra reference would change the task.

The graph contains a named editing subgraph. To see its nodes, double-click its empty area or use its edit button; this feature requires frontend 1.24.3 or later. For this non-Lightning baseline, right-click the internal LoraLoaderModelOnly node and set its mode to Bypass (pass the model through without loading that optional adaptation). Leave turbo false. This avoids keeping an active LoRA loader pointed at a file you have not downloaded. Missing core nodes require checking the documented compatible update, rather than installing arbitrary custom packs. This adaptation is inspected guidance, not an executed graph.

Generate, compare and export

1. Replace positive_prompt with the following starter. Keep the negative prompt empty for this baseline, and retain the non-turbo sampling branch's settings. Record the seed actually used.

Use image 1 as the character reference. Place the same cream robot
standing in a small kitchen. Preserve its rectangular cream head,
two dark circular eyes, body proportions and orange jacket.
Change the surroundings only. Show one character, with no text.

2. Run one job with Queue. The prompt separates the requested scene change from the traits you will inspect; it does not lock those traits mechanically.
3. Inspect head silhouette, eyes, jacket and proportions against the original. Reject an attractive scene if those required features have changed. Save each accepted PNG from the SaveImage output; the default local output directory is ComfyUI/output unless you changed it.
4. Save the modified workflow JSON and a log containing ComfyUI version, component filenames, seed, settings, input, elapsed time and acceptance notes. Repeat from the original reference, changing only the kitchen instruction for the other two scenes.

Paid browser route: Midjourney Edit Model

Use the current Edit Model instructions, which document V8.1/V8.2 compatibility. Older V7 Omni Reference instructions use different controls.

1. Sign in and review your plan, remaining allowance and image visibility. In the Imagine bar, open the image panel and upload the original reference.
2. Attach it to the prompt, select a compatible V8 version, and enter the same scene-and-identity instruction. Keep a single reference for the first comparison.
3. Generate and inspect the same four traits. Repeat from the original reference for each scene, rather than assuming a variation retains identity.
4. Open an accepted image in the Editor when you need a targeted repair. Use its export/download control for the generated image; keep the reference and job details with it.

When the character changes

If the face drifts, reduce the requested change and test one scene with the same input. If reference clothing appears where you wanted different clothing, state which traits may change; identity and wardrobe are separate requirements. If several people blend together, return to one subject before attempting a group.

A LoRA may help some repeatable production tasks, but training is a separate project with data, compatibility and evaluation requirements. Establish the reference-editing baseline first. No result image, reliability percentage or execution benchmark is supplied by this guide.

Use image 1 as the character reference. Place the same cream robot standing in a small kitchen. Preserve its rectangular cream head, two dark circular eyes, body proportions and orange jacket. Change the surroundings only. Show one character, with no text.